Method and device for predicting willingness to pay for user, storage medium and electronic equipment

Abstract

The invention belongs to the technical field of freight rate search, and provides a user payment willingness prediction method and device, a storage medium and electronic equipment, and the method comprises the steps: generating corresponding session representations based on the feature representations of a plurality of flight projects; evolution is carried out on preferences of different time levels of the target user on the basis of a session representation and feature representation evolution network, and preference feature representation of the target user is obtained; based on the preference feature representation and the probability distribution of the target user, determining a willingness-to-pay probability of the target user for a target flight in all flights; and based on the willingness-to-pay probability of the target flight, predicting the willingness-to-pay of the target user for the target flight to obtain a prediction result. According to the prediction method, the feature evolution mechanism of the target user is introduced, so that the flight information interacting with the target user is reserved, the obtained preference features of the target user are more comprehensive, the finally obtained prediction result is more accurate, and the prediction result better conforms to the actual situation of the target user.

Description

Technical Field

[0001] The present invention belongs to the technical field of fare search, and in particular relates to a method, device, storage medium and electronic device for predicting user willingness to pay. Background Art

[0002] The user willingness to pay forecast is based on different data processing points before the flight departs, and predicts the future user flight demand and willingness to pay, which serves as the basis for controlling the number of flight seats and prices. The concept of willingness to pay comes from the field of economics. The meaning of willingness to pay is: the highest price that individual consumers are willing to pay for goods or services. In the field of revenue management, the most effective way to increase airline revenue is to set the willingness to pay price that different customers can accept.

[0003] Early research on willingness to pay was mainly reflected in the customer's high-buy rate. High-buy is defined as the behavior of customer demand transferring from a closed cabin to a higher-priced cabin. Researchers have improved the traditional EMSRb algorithm and proposed a new heuristic algorithm. Another study simulated the algorithm in the PODS (Passenger Origin-Destination Simulator) system, but did not provide a method to estimate the customer's high-buy rate. Based on the concept of high-buy, the Q-forecasting prediction algorithm was proposed to estimate the demand for those who are only willing to buy the lowest cabin (called Q-class) and the probability of these demands buying up to a higher cabin. Further research defined these two different demands as price-oriented demand and product-oriented demand, and designed the HF (Hybrid forecasting) method to estimate and predict these two demands separately. There are also studies that use linear regression to characterize customers' high-buy behavior and use it to predict the booking demand for each cabin.

[0004] The willingness to pay of these methods is mainly reflected in the characterization of demand functions. For example, the customer's demand function for a given price can be expressed as the probability that the customer is willing to pay no less than the given price. In order to solve the problem of freedom in the estimation of demand parameters, some studies have introduced the external variable of market share to limit the freedom between the arrival rate and the MNL (Multinomial Logit) model demand parameters. Another study split the simultaneous estimation task of the arrival rate and the selection of model demand parameters into a two-step estimation task. For the estimation problem of customer demand models, some studies have given the conditions for the uniqueness and identifiability of MNL model parameter estimation, and proposed an MM (Minorization-maximization) demand estimation algorithm. The above studies are all based on the MLE (Maximum likelihood estimation) method to model the parameter estimation task, and some scholars have also chosen other parameter estimation methods to characterize the demand parameter estimation tasks under different scenarios. For example, a study proposed a LM (Loss Minimization) algorithm based on MIP (Mixed-Integer Programming) to simultaneously solve the customer arrival rate and MNL model parameters. Another study took into account the endogeneity problem between the civil aviation company's own cabin control behavior and the observed sales data and proposed a two-step generalized moment estimation method. Based on the EM (Expectation Maximization) algorithm, a study proposed an estimation method for the Markov chain selection model. Another study characterized the estimation task as a constrained convex optimization problem and applied the Frank-Wolfe algorithm to solve it. Another study proposed a new dynamic pricing strategy. The difference between this method and other methods is that it decomposes demand to the level of individual customers. Each customer has a different willingness to pay, and the willingness to pay of customer groups forms a distribution of customer group willingness to pay.

[0005] In recent years, domestic scholars have also begun to predict users' willingness to pay based on the actual operating data of airlines, and this is mainly reflected in the application of user selection models to the dynamic pricing of airlines. Some studies use non-homogeneous Poisson processes to simulate user arrival and use discrete choice models to quantify user purchase willingness and probability. Other studies use the Logit model to predict users' willingness to buy based on user arrival and user market segmentation. In addition, researchers use expected marginal benefits, equilibrium perspectives, and customer perceived value to perform revenue optimization analysis on the original high-buy rate model. However, the key user selection models in the above studies are all developed from the perspective of airlines. Some studies start from the user's perspective, based on user attributes and multi-behavior feedback, use personalized recommendation algorithms to predict users' willingness to pay and conduct dynamic pricing research.

[0006] Whether it is the characterization of user demand functions, the estimation of demand parameters, or the construction of user selection models, the prediction of willingness to pay is all about controlling the number of seats and prices from the perspective of airlines, and the willingness to pay cannot be decomposed to each user level. Although some studies have decomposed demand to the level of individual users, they are still estimated from the perspective of all user behaviors. Some studies have estimated the willingness to pay of each user from the user's perspective by using the selection behavior history of each user and personalized recommendations, but they have not considered the evolution of user preferences at different time levels. In practice, user preferences change with time periods, and each user has different preferences for time.

[0007] How to accurately predict users' willingness to pay is a technical problem to be solved. Summary of the invention

[0008] Based on this, it is necessary to provide a method, device, storage medium and electronic device for predicting user willingness to pay to address the defect that existing prediction methods cannot accurately predict user willingness to pay.

[0009] In a first aspect, an embodiment of the present invention provides a method for predicting a user's willingness to pay, the method comprising:

[0010] Based on the continuous numerical features and discrete categorical features in the flight feature information, generate feature representations of multiple flight items corresponding to the flight feature layer;

[0011] generating corresponding conversation representations based on the feature representations of the plurality of flight items;

[0012] Based on the session representation and feature representation evolution network, the preferences of the target user at different time levels are evolved to obtain the preference feature representation of the target user;

[0013] Determining the target user's willingness to pay probability for the target flight among all flights based on the target user's preference characteristic representation and probability distribution;

[0014] Based on the willingness-to-pay probability of the target flight, the willingness-to-pay of the target user for the target flight is predicted to obtain a prediction result.

[0015] Optionally, determining the probability of the target user's willingness to pay for the target flight among all the flights based on the preference feature representation and probability distribution of the target user includes:

[0016] By means of a polynomial decoder, the preference characteristic representation of the target user is mapped into a probability distribution of the target user's willingness to pay for all flights;

[0017] Obtaining a probability distribution, and using the probability distribution as a target distribution to which the willingness-to-pay probability distribution is approximated;

[0018] The willingness to pay probability distribution and the probability distribution are minimized and optimized by the KL divergence method to obtain the willingness to pay probability of the target flight.

[0019] Optionally, before obtaining the probability distribution, the method further includes:

[0020] The target flight is converted into a flight number vector for the target flight in a preset dimension by using a one-hot encoding method; and the flight number vector is used as the probability distribution.

[0021] Optionally, the evolving network based on the session representation and feature representation to evolve preferences of target users at different time levels includes:

[0022] Acquire the session representation, where the session representation is used to replace the interaction process between all flights in the session and the target user;

[0023] For the target user, the update function of the feature representation evolution network is used to express the target user and the kth session The feature evolution process when interaction occurs is to obtain the feature representation generated by the kth interaction and the preference representation generated by the k-1th interaction, where u j For the target user, the long-term features and short-term change features of the target user are aggregated through the feature evolution process, and the neural networks used to aggregate the long-term features and the short-term change features include: RNN, LSTM and GRU.

[0024] Optionally, the generating corresponding session representations based on the feature representations of the plurality of flight items includes:

[0025] Get flight feature sequence;

[0026] Arranging the flight feature sequences in sequence according to the chronological order of the time-stamped interaction events;

[0027] Obtaining feature representations of the plurality of flight items;

[0028] The feature representations of the plurality of flight items are mapped to corresponding session representations through a preset aggregation function, and the neural network adopted by the preset aggregation function is LSTM.

[0029] Optionally, the generating of feature representations of a plurality of flight items corresponding to the flight feature layer based on the continuous numerical features and the discrete category features in the flight feature information includes:

[0030] Acquire flight feature information, the flight feature information including information corresponding to the continuous numerical feature and information corresponding to the discrete category feature, the continuous numerical feature is a dense feature, and the discrete category feature is a sparse feature;

[0031] For continuous numerical features, the numerical features are projected and converted into a first feature representation through a projection matrix, where the first feature representation includes a projection matrix and a bias matrix;

[0032] For discrete category features, numerical processing is performed on the discrete category features to convert the category identifier into a number;

[0033] The discrete category features processed by numerical processing are converted to obtain dense low-dimensional features; the numbers are used as indexes to obtain the second feature representation by looking up the table;

[0034] The first feature representation and the second feature representation are obtained, and the first feature representation and the second feature representation are concatenated to generate feature representations of a plurality of flight items corresponding to the flight feature layer.

[0035] Optionally, before generating feature representations of a plurality of flight items corresponding to the flight feature layer based on the continuous numerical features and discrete category features in the flight feature information, the method further includes:

[0036] Acquire an event sequence, and divide the event sequence into the session set;

[0037] For the session set, obtaining the number of sessions retained by the target user; and obtaining the length of events retained in one session;

[0038] Obtain a time window T, and use the time window T as a session division mark;

[0039] When the time interval between two adjacent events in the event sequence is greater than the time window T, the two adjacent events are divided into different sessions to divide the session of the target user;

[0040] The kth session S of the target user k The time when the first interaction event in the set occurs is taken as the session S k Time mark;

[0041] A time stamp set of the session is set, wherein the time stamp set includes time information at the hour level, time information at the week level, and time information at the month level.

[0042] In a second aspect, an embodiment of the present invention provides a device for predicting a user's willingness to pay, the device comprising:

[0043] A first generation module, used to generate feature representations of multiple flight items corresponding to the flight feature layer based on continuous numerical features and discrete category features in the flight feature information;

[0044] A second generating module, configured to generate corresponding session representations based on the feature representations of the plurality of flight items;

[0045] An evolution module, used to evolve the preferences of the target user at different time levels based on the session representation and feature representation evolution network to obtain the preference feature representation of the target user;

[0046] A determination module, configured to determine the probability of the target user's willingness to pay for a target flight among all flights based on the target user's preference feature representation and probability distribution;

[0047] The prediction module is used to predict the payment willingness of the target user for the target flight based on the payment willingness probability of the target flight to obtain a prediction result.

[0048] According to a third aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute the method of the first aspect.

[0049] According to a fourth aspect, an electronic device is provided, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method according to the first aspect is implemented.

[0050] In an embodiment of the present invention, based on the continuous numerical features and discrete category features in the flight feature information, feature representations of multiple flight items corresponding to the flight feature layer are generated; based on the feature representations of multiple flight items, corresponding session representations are generated; based on the session representation and feature representation evolution network, the preferences of the target user at different time levels are evolved to obtain the target user's preference feature representation; based on the target user's preference feature representation and probability distribution, the target user's willingness to pay probability for the target flight among all flights is determined; and based on the target flight's willingness to pay probability, the target user's willingness to pay for the target flight is predicted. The user willingness to pay prediction method provided by the embodiment of the present invention introduces a feature evolution mechanism of the target user, so that the flight information interacting with the target user can be retained, the obtained target user's preference features are more comprehensive, and the final prediction result is more accurate, and the prediction result is more in line with the actual situation of the target user. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] By referring to the following drawings, a more complete understanding of the exemplary embodiments of the present invention can be obtained. The drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the drawings, the same reference numerals generally represent the same components or steps.

[0052] Figure 1 A flowchart of a method for predicting user willingness to pay according to an exemplary embodiment of the present invention;

[0053] Figure 2 A flowchart for generating a feature representation of user preferences based on time evolution;

[0054] Figure 3 It is a model diagram of temporal feature extraction and feature evolution;

[0055] Figure 4 For payment intention prediction model diagram;

[0056] Figure 5 It is a schematic diagram of the structure of a device 500 for predicting user willingness to pay according to an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0057] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0058] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in the present invention should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0059] In addition, the terms "first" and "second" etc. are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, product or equipment that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or equipment.

[0060] Embodiments of the present invention provide a method and device for predicting user willingness to pay, a computer-readable medium, and an electronic device, which are described below in conjunction with the accompanying drawings.

[0061] Please refer to Figure 1 , which shows a flow chart of a method for predicting user willingness to pay provided by some embodiments of the present invention, such as Figure 1 As shown, the method for predicting user willingness to pay may include the following steps:

[0062] Step S101: generating feature representations of a plurality of flight items corresponding to a flight feature layer based on continuous numerical features and discrete category features in flight feature information.

[0063] In one example, based on continuous numerical features and discrete category features in flight feature information, generating feature representations of multiple flight items corresponding to a flight feature layer includes the following steps:

[0064] Acquire flight feature information, which includes information corresponding to continuous numerical features and information corresponding to discrete category features. Continuous numerical features are dense features, and discrete category features are sparse features.

[0065] For continuous numerical features, the numerical features are projected and converted into a first feature representation through a projection matrix. The first feature representation includes a projection matrix and a bias matrix.

[0066] For discrete category features, the discrete category features are numerically processed to convert the category identifiers into numbers;

[0067] The discrete category features that have been numerically processed are converted to obtain dense low-dimensional features; the second feature representation is obtained by using numbers as indexes and looking up tables;

[0068] The first feature representation and the second feature representation are obtained, and the first feature representation and the second feature representation are concatenated to generate feature representations of multiple flight items corresponding to the flight feature layer.

[0069] In one example, the characteristic information of the flight Contains two types of features, |i| is the number of flight features, and one is a continuous numerical feature It is a dense feature data; the other is a discrete category feature It is sparse feature data. |i| d and |i| s are the number of features of dense features and sparse features, respectively, and |i| d +|i| s =|i|.

[0070] For continuous features, the projection matrix is ​​used to transform the numerical feature projection into the corresponding feature representation:

[0071] emb den =σ(I den W den +B den )

[0072] in, are the trainable projection matrix and bias matrix respectively, and σ(.) is the activation function ReLU.

[0073] For discrete data features, first digitize the discrete features and convert the category identifiers into numbers. Then convert them into dense low-dimensional features. Use numbers as indexes and obtain their feature representation by looking up a table.

[0074] The feature representation of the flight item is composed of dense feature representation and sparse feature representation:

[0075]

[0076] in, For splicing operation, h i is the dimension of flight feature representation, h i =h d +h s ,emb item is the feature representation of the flight item, emb dense is a dense feature representation, emb sparse It is a sparse feature representation.

[0077] In one example, before generating feature representations of multiple flight items corresponding to the flight feature layer based on the continuous numerical features and discrete category features in the flight feature information, the method for predicting user willingness to pay provided by the embodiment of the present invention may further include the following steps:

[0078] The sessions of the target users are divided to obtain a session division layer.

[0079] In one example, dividing the sessions of the target user includes the following steps:

[0080] Obtain an event sequence and divide the event sequence into a session set;

[0081] For the session set, obtain the number of sessions retained by the target user; and obtain the length of events retained in a session;

[0082] Obtain a time window T, and use the time window T as a session division mark;

[0083] When the time interval between two adjacent events in the event sequence is greater than the time window T, the two adjacent events are divided into different sessions to divide the session of the target user;

[0084] The kth session S of the target user k The time when the first interaction event in the set occurs is taken as the session S k Time mark;

[0085] Set the time stamp set of the session, which includes time information at the hour level, week level, and month level.

[0086] In one example, for an event sequence Q = {e1, e2, ..., e n}, and divide it into a session set S = {S1, S2, …, S |S|}, |S| is the number of sessions reserved for the target user. The kth session S of the target user u k Indicated as S k ={e k1 ,e k2 ,…,e k|I|}, where |I| is the length of events retained in a session. Set the time window T as the division mark of the session. When the time interval between two adjacent events in Q is greater than T, the two events are divided into different sessions. k The first interaction event in the collection k1 The occurrence time t sess As a session kSet the session time stamp set unisSet, which contains time information at three levels: hour, week, and month. h ={1,2,…,24},unit w ={1,2,…,7},unit m ={1,2,…,12}.

[0087] In one example, the method for predicting user willingness to pay provided by the embodiment of the present invention may further include the following steps:

[0088] In the session graph, if the target user interacts with a flight in a session, it is determined that there is a common edge between the session node and the flight node.

[0089] In one example, in a session graph, if a target user interacts with a flight node in a session, determining that there is a common edge between a session node and a flight node includes the following steps:

[0090] For the target user, a subset of the target user's session set, a subset of the target user's flight set, and an interaction event set of the target user are obtained;

[0091] Constructing a conversation graph based on a subset of the target user's conversation set, a subset of the target user's flight set, and a subset of the target user's flight set and a set of interaction events of the target user, wherein the set of interaction events of the target user also includes time information of each interaction event;

[0092] In the session graph, if the target user interacts with a flight in a session, it is determined that there is a common edge between the session node and the flight node.

[0093] In one example, for the target user u∈U, the subset S of the conversation set u ∈S, a subset of the flight set F u ∈F. The set of interaction events E u Record the occurrence of each interaction event. Construct the conversation graph G u ={S u ,F u ,E u}.

[0094] In G u In session S, if the target user u is in session S i In the interaction with the flight, S i Node and F j There is an edge (S i ,F j )∈E u , while E u It also contains the time information of each interaction event tij The interaction event set E describes a many-to-many relationship.

[0095] Step S102: Generate corresponding conversation representations based on the feature representations of multiple flight items.

[0096] In one example, generating corresponding conversation representations based on feature representations of multiple flight items includes the following steps:

[0097] Get flight feature sequence;

[0098] Arrange the flight feature sequences in order according to the chronological order of the time-stamped interaction events;

[0099] Obtain feature representations of multiple flight items;

[0100] Through the preset aggregation function, the feature representations of multiple flight items are mapped to corresponding session representations, and the neural network used by the preset aggregation function is LSTM.

[0101] In one example, based on the feature representation of the flight item, the flight feature sequence is arranged in the chronological order of the time-stamped interaction events, that is, S k =(emb item,1 ,emb item,2 ,…,emb item,n ), n is the maximum number of flight items in a session, and the feature representations of multiple flight items are mapped to a session representation through an aggregation function:

[0102]

[0103] Among them, σ(.) is the activation function, is a trainable nonlinear projection matrix, is a trainable bias matrix. sesS (.) Aggregation function, whose input vector dimension is the dimension size of the flight item feature item , the output is the session representation dimension h sess There are many choices for aggregation functions, such as RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), multi-head attention network, etc. Considering that the flight items in the session are organized according to the interaction events t j,i For arrangement, a long short-term memory network that can integrate contextual information is selected as the aggregation function.

[0104] Step S103: Based on the session representation and feature representation evolution network, the preferences of the target user at different time levels are evolved to obtain the preference feature representation of the target user.

[0105] In one example, based on the session representation and feature representation evolution network, the preferences of target users at different time levels are evolved, including the following steps:

[0106] Get the session representation, which is used to represent the interaction process between all flights and the target user in the session;

[0107] For the target user, the update function of the feature representation evolution network is used to express the target user and the kth session The feature evolution process when interaction occurs is to obtain the feature representation generated by the kth interaction and the preference representation generated by the k-1th interaction, where u j For the target user, the feature evolution process is used to aggregate the long-term features and short-term change features of the target user. The neural networks used to aggregate the long-term features and short-term change features include: RNN, LSTM and GRU.

[0108] In one example, for the target user u j , an update function is used to describe the relationship between the target user and the kth session The process of feature evolution when interaction occurs:

[0109]

[0110] represents the target user u j The feature representation after the kth interaction occurs, represents the preference expression generated in the k-1th interaction. f(·) is the interaction change function, which is used to calculate The new impact of the change in the target user's preference. Agg(·) represents the aggregation function used to aggregate the historical preferences and updated parts of the target user's preferences. The essence of this process is to aggregate the long-term features and short-term change features of the target user. The aggregation process can be implemented using neural networks such as RNN, LSTM, GRU (Gated Recurrent Unit, gated recurrent unit).

[0111] Step S104: Determine the target user's willingness to pay probability for the target flight among all flights based on the target user's preference feature representation and probability distribution.

[0112] In one example, determining the probability of the target user's willingness to pay for a target flight among all flights based on the target user's preference feature representation and probability distribution includes the following steps:

[0113] Through the polynomial decoder, the target user's preference feature representation is mapped into the target user's willingness to pay probability distribution for all flights;

[0114] Obtain a probability distribution, and use the probability distribution as a target distribution to approximate the probability distribution of willingness to pay;

[0115] Through the KL divergence method, the willingness to pay probability distribution and probability distribution are minimized and optimized to obtain the willingness to pay probability of the target flight.

[0116] In one example, a polynomial decoder is used to map the target user's preference feature representation to the target user's willingness to pay probability distribution for all flights through a nonlinear mapping:

[0117] π(emb u )=Softmax(σ(W d emb u +B d ))

[0118] Among them, W d is the trainable projection matrix, B d is the trainable bias matrix.

[0119] The target user’s final interactive flight is converted into a vector with a dimension equal to the number of flights by one-hot encoding, and used as a probability distribution Q, which represents the probability distribution of the target user’s willingness to pay for all flights, which is π(emb u ) The target distribution to be approximated is minimized by KL divergence. u ) and Q, and use them as optimization objectives:

[0120]

[0121] Among them, M is the number of target user tables, g is the number of flights, For target user u j The probability of willingness to pay for flight k.

[0122] In one example, before obtaining the probability distribution, the method for predicting user willingness to pay provided by the embodiment of the present invention may further include the following steps:

[0123] Using one-hot encoding, the target flight is converted into a flight number vector for the target flight in a preset dimension; and the flight number vector is used as a probability distribution.

[0124] Step S105: Based on the willingness-to-pay probability of the target flight, the willingness-to-pay of the target user for the target flight is predicted to obtain a prediction result.

[0125] In actual application scenarios, if the probability of willingness to pay for the target flight is higher, the prediction result is: the target user is more willing to pay for the target flight, and the target flight can be recommended to the target user; conversely, if the probability of willingness to pay for the target flight is lower, the prediction result is: the target user is less willing to pay for the target flight.

[0126] In actual application scenarios, users have very obvious time periodicity when flying, and this time periodicity is reflected in multiple time levels. For example, from the perspective of passenger flow, the passenger flow in the early morning and midnight of a day is less than that in the morning and afternoon; the passenger flow on weekdays is less than that near the weekend; in a year, the passenger flow during the golden week of various holidays and winter and summer vacations will be significantly more than the rest of the day. This change in passenger flow directly leads to changes in flight fares, which is specifically manifested in that the ticket price of the same route during peak traffic periods will be higher. This time periodicity reflects that users have different flight preferences at different times and have their own characteristics in different time dimensions. For example, users tend to choose expensive business class on weekdays due to business needs, and choose cheaper economy class on non-working days; some users tend to choose cheap midnight tickets, while others are less sensitive to prices and tend to choose daytime flights.

[0127] Capturing the target user's time-periodic flight preference is crucial for predicting willingness to pay. Therefore, it is necessary to discover and evolve the characteristics of the target user at different time levels (hour, week, month), obtain the target user's dynamic preference, and design a personalized recommendation algorithm to improve the accuracy of the prediction of willingness to pay.

[0128] In one example, the method for predicting user willingness to pay may include the following steps:

[0129] Step a1: Generate a feature representation of target user preferences based on time evolution.

[0130] The process of generating the feature representation of target user preferences based on time evolution is as follows: Figure 2 As shown, feature representations are extracted from flight and user dimensions, aggregation and preference evolution are performed in different time dimensions, and user willingness to pay is predicted to obtain prediction results.

[0131] like Figure 2 As shown in Figure 2, the process of generating the target user preference feature representation based on time evolution is as follows:

[0132] Step a11: Divide the sessions of target user behavior.

[0133] The target user's preferences will change over time and with the occurrence of interactive events. Therefore, the target user's interaction history is divided into certain time intervals. The preference consistency within a session is shown in a short period of time; the dynamic differences in the target user's preferences are shown between sessions. The target user's multiple session records are not only different, but also connected.

[0134] For the target user u∈U, the subset S of the conversation set u ∈S, a subset of the flight set F u ∈F. The set of interaction events E u Record the occurrence of each interaction event. Construct the conversation graph G u ={S u ,F u ,E u}.

[0135] In G u In session S, if the target user u is in session S i In the interaction with the flight, S i Node and F j There is an edge (S i ,F j )∈E u , while E u It also contains the time information of each interaction event t ij The interaction event set E describes a many-to-many relationship.

[0136] For the event sequence Q = {w1, e2, ..., e n}, and divide it into a session set S = {S1, S2, …, S |S|}, |S| is the number of sessions reserved for the target user. The kth session S of the target user u k Indicated as S k ={e k1 ,e k2 ,…,e k|I|}, where |I| is the length of events retained in a session. Set the time window T as the division mark of the session. When the time interval between two adjacent events in Q is greater than T, the two events are divided into different sessions. k The first interaction event in the collection k1 The occurrence time t sess As a session k Set the time stamp set UnitSet of the session, which contains time information at three levels: hour, week, and month. h ={1,2,…,24},unit w ={1,2,…,7},unitm ={1,2,…,12}.

[0137] Step a12: The dense feature representation and the sparse feature representation together constitute the feature representation of the flight item.

[0138] Flight characteristics information Contains two types of features, |i| is the number of flight features, and one is a continuous numerical feature It is a dense feature data; the other is a discrete category feature It is sparse feature data. |i| d and |i| s are the number of features of dense features and sparse features, respectively, and |i| d +|i| s =|i|.

[0139] For continuous features, the projection matrix is ​​used to transform the numerical feature projection into the corresponding feature representation:

[0140] emb den =σ(I den W den +B den )

[0141] in, are the trainable projection matrix and bias matrix respectively, and σ(.) is the activation function ReLU.

[0142] For discrete data features, first digitize the discrete features and convert the category identifiers into numbers. Then convert them into dense low-dimensional features. Use numbers as indexes and obtain their feature representation by looking up a table.

[0143] The feature representation of the flight item is composed of dense feature representation and sparse feature representation:

[0144]

[0145] in, For splicing operation, h i is the dimension of flight feature representation, h i =h d +h s ,emb item is the feature representation of the flight item, emb dense is a dense feature representation, emb sparse It is a sparse feature representation.

[0146] Step a13: Generate preference feature representations of target users at different time levels.

[0147] On the basis of obtaining the feature representation of the flight items, the flight feature sequences are arranged in the chronological order of the time-stamped interaction events, that is, S k =(emb item,1 ,emb item,2 ,…,emb item,n ), n is the maximum number of flight items in a session, and the feature representations of multiple flight items are mapped to a session representation through an aggregation function:

[0148]

[0149] Among them, σ(.) is the activation function, is a trainable nonlinear projection matrix, is a trainable bias matrix. sess (.) Aggregation function, whose input vector dimension is the dimension size of the flight item feature item , the output is the session representation dimension h sess There are many choices for aggregation functions, such as RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), multi-head attention network, etc. Considering that the flight items in the session are organized according to the interaction events t j,i For arrangement, a long short-term memory network that can integrate contextual information is selected as the aggregation function.

[0150] Step a14: Evolving the preference feature representations of the target user at different time levels to generate a preference feature representation of the target user based on time evolution.

[0151] The target user's preferences are not always the same. As the target user interacts with the item, the target user will often have new preferences as the interaction occurs. In order to capture this change, a feature representation evolution network is used, such as Figure 3 As shown. Figure 3 As shown in the figure, the feature representation of each link can be aggregated and evolved using algorithms such as GRU (Gated Recurrent Unit) and RNN. The session representation is used to replace all flights in the session to interact with the target user. In this way, on the one hand, compared with the interaction and evolution between the target user and all flights, the calculation amount of the model can be reduced and the calculation speed can be accelerated; on the other hand, the browsing events of the target user are integrated with the purchase events, which solves the data cold start problem caused by the sparse purchase events.

[0152] For target user u j, an update function is used to describe the relationship between the target user and the kth session The process of feature evolution when interaction occurs:

[0153]

[0154] represents the target user u j The feature representation after the kth interaction occurs, represents the preference expression generated in the k-1th interaction. f(·) is the interaction change function, which is used to calculate The new impact of the change in the target user's preference. Agg(·) represents the aggregation function used to aggregate the historical preferences and updated parts of the target user's preferences. The essence of this process is to aggregate the long-term characteristics and short-term change characteristics of the target user. The aggregation process can be implemented using neural networks such as RNN, LSTM, and GRU.

[0155] Step a2: After obtaining the target user preference feature representation based on time evolution, predict the user's willingness to pay. The process of predicting the user's willingness to pay is as follows:

[0156] The target user's preference feature representation is mapped to the target user's willingness to pay probability distribution for all flights through a polynomial decoder through a nonlinear mapping:

[0157] π(emb u )=Softmax(σ(W d emb u +B d ))

[0158] Among them, W d is the trainable projection matrix, B d is the trainable bias matrix.

[0159] The target user’s final interactive flight is converted into a vector with a dimension equal to the number of flights using a one-hot encoding method, and is used as a probability distribution Q to represent the target user’s willingness to pay for all flights. This is π(emb u ) The target distribution to be approximated is minimized by KL divergence. u ) and Q, and use them as optimization objectives:

[0160]

[0161] Among them, M is the number of target user tables, g is the number of flights, For target user u j For the willingness to pay probability of flight k, the specific process is as follows Figure 4 As shown. Figure 4 As shown, the preference feature representation of the target user and the clustered target user are combined and the payment willingness of the target user for different discounted flights is predicted.

[0162] The method for predicting user willingness to pay provided by the embodiment of the present invention utilizes the user's search and interaction history to capture the user's preferences and characteristics in different time periods, and utilizes two neural networks to evolve the user's preferences and user characteristics, thereby obtaining the user's evolved user preferences and predicting the willingness to pay. Compared with the traditional method, it can not only improve the prediction accuracy, but also effectively meet the different flight preferences of users in different time periods.

[0163] In another example, the method for predicting user willingness to pay may include the following steps:

[0164] Step b1: Acquire usage data and obtain the data format.

[0165] Step b11: Obtain usage data. The behavioral data of approximately 28,000 users on 202 routes of a domestic Chinese airline from September 1, 2020 to November 30, 2020 are used, totaling 1.048 million behavioral events. The difference is that in order to capture the characteristic preferences of users at different time levels, the time when the user interaction events occurred is added, and the training set and test set are divided by time. The user interaction data from September 1, 2020 to November 15, 2020 is used as the training set data; the user interaction data from November 16, 2020 to November 30, 2020 is used as the test set data.

[0166] Step b12: Obtain data format.

[0167] Training data table (SX_train_dy.txt): Each row represents an interaction record between a user and an item, and the column names are: uid (user ID), iid (item ID), t (time), label (event type).

[0168] Test data table (SX_test_dy.txt): Column names: uid (user ID), iid (project ID), t (time), label (event type); each row represents a record of interaction between a user and a project.

[0169] User attribute table (attr_dy.txt): Column names: uid (user ID), attr1, attr2, attr3, attr4, attr5, attr6 (user attributes); each row represents the attribute information of the user.

[0170] Step b2: Load training data.

[0171] Input parameters: file: data file path, i.e. . / data / SX_train_dy.txt.

[0172] Data loading: Use the pd.read_csv function to load the data file, use the tab character ('\t') as the delimiter, and specify the column names as ['uid', 'iid', 't', 'label']. The column names here represent user ID, project ID, timestamp, and label. Organization of time windows: Use groupby('t') to group the data by timestamp (t). This step divides the data into multiple groups based on timestamps, and each group represents a time window.

[0173] Organize data: By traversing the time window group (antDict), extract the user ID, project ID, and label in each time window into user, item, and lab lists respectively. Store these lists in antlist to form a three-layer nested list.

[0174] Return value: antlist: A list containing three sublists, storing user IDs, project IDs and tags respectively. Each sublist represents a user, project and tag within a time window.

[0175] Return result example:

[0176] For a certain time window, the structure of antlist may be as follows: [

[0178] [u1,u2,...],#User ID list

[0179] [i1,i2,...],#project ID list

[0180] [l1,l2,...]#label list ]

[0182] Among them, u1, i1, l1 represent a user interaction event, u2, i2, l2 represent another user interaction event, and so on.

[0183] Step b3: Preprocess the data.

[0184] Read two data files named user_record_discount.csv and user_info_discount.csv, which contain user records and user attribute information.

[0185] Convert the time field in the record to a timestamp and divide the data into a training set and a test set based on the timestamp. The timestamp threshold is '2020 / 11 / 16 00:00:00'.

[0186] Traverse the user records in the training set and extract the user's attribute information, including sparse features and dense features. At the same time, organize the user's events into sessions in chronological order and generate a user session list and session time list.

[0187] In the formed session, the last shopping event is found according to the event flag (1 indicates a shopping event) as a candidate item for each user in the training set, and the time of the candidate item is recorded.

[0188] Convert the time information in the training set to a list of months, days of the week, and hours.

[0189] Record the session length for each user in the training set.

[0190] Similarly, traverse the user records of the test set, extract the user's attribute information, generate a list of user sessions and a list of session times, find the candidate items for each user in the test set, and record the time of the candidate items.

[0191] Convert the time information in the test set to a list of months, days of the week, and hours.

[0192] Record the session length for each user in the test set.

[0193] Returns all generated information.

[0194] Step b4: Obtain preprocessed data.

[0195] In one example, the user-related data specifically includes:

[0196] 1.user_dict_train,user_dict_test: Dictionaries containing information such as training and test user attributes, target items, user sessions, session time, and session length.

[0197] {user_id:[user_attr,target_item,user_sessions,session_times,session_lengths]}

[0198] 2.user_setime_train,user_setime_test: user session time information.

[0199] {user_id:[session1_time,session2_time,...]}

[0200] 3.user_sesslength_train,user_sesslength_test: user’s session length information.

[0201] {user_id:[session1_length,session2_length,...]}

[0202] In one example, the candidate project related data specifically includes:

[0203] train_can,test_can: Dictionary containing candidate items for training and testing users.

[0204] {user_id:candidate_item}.

[0205] Step b5: Generate user embeddings at different time levels.

[0206] Step b51: Generate user and item representations.

[0207] The user index is obtained through user_embeds and stored in user_embed. The candidate item index is obtained through item_embed and stored in candidate_embed.

[0208] Step b52: Generate session representation.

[0209] The item sequence in input_sess is obtained through item_embed to get the embedded representation of the item and stored in sess_embed. sess_embed is converted to session embedding using item2sess and stored in sess_embed.

[0210] Step b53: Extract time features.

[0211] The time features of the three time levels (hour, week, month) are extracted by the get_timeembed function and stored in time_feas.

[0212] Step b54: Update user representation.

[0213] Update the corresponding user representation in user_embed_update to time_feas.

[0214] Step b55: Activate and merge.

[0215] Concatenate the time feature time_feas and the session embedding sess_embed.

[0216] Step b56: Calculate the score.

[0217] The scores are calculated through fully connected layers (score_fc1 and score_fc2).

[0218] Step b57: Calculate Loss.

[0219] During the training phase, the loss is calculated using LogSoftmax and Negative LogLikelihood (NLL).

[0220] Negative sampling: During the training phase, negative sampling is performed to calculate the fraction of negative samples.

[0221] Output: Returns the calculated loss and the predicted probability or score for a given user and candidate item.

[0222] Step b6: Evolve the preference feature representations of users at different time levels.

[0223] Enter the session representation emb sess , user feature representationemb user , time level time_level, time level unit set UnitSet, session time t sess ;

[0224] The conversation sequence is divided into different time units, aggregated to obtain the conversation sequence representation, and GRUCELL is used to evolve the preference feature representation of users at different time levels.

[0225] Step b7: Load test data.

[0226] Input parameters: file: data file path, i.e. . / data / SX_test_dy.txt.

[0227] Data loading: Use the pd.read_csv function to load the data file, use the tab character ('\t') as the delimiter, and specify the column names as ['uid', 'iid', 't', 'label']. The column names here represent user ID, project ID, timestamp, and label.

[0228] Organization of time windows:

[0229] Use groupby('t') to group the data by timestamp(t). This step divides the data into multiple groups based on timestamps, each group represents a time window.

[0230] Organizing data:

[0231] By traversing the time window group (testDict), the user ID, project ID, and label in each time window are extracted into user, item, and lab lists respectively. These lists are stored in testlist to form a three-layer nested list.

[0232] Return value:

[0233] testlist: A list containing three sublists, storing user IDs, project IDs, and tags respectively. Each sublist represents a user, project, and tag within a time window.

[0234] Return result example:

[0235] For a certain time window, the structure of testlist may be as follows: [

[0237] [u1,u2,...],#User ID list

[0238] [i1,i2,...],#project ID list

[0239] [l1,l2,...]#label list

[0240] ]Among them, u1, i1, l1 represent a user interaction event, u2, i2, l2 represent another user interaction event, and so on.

[0241] In general, the main function of this function is to load the test data file, organize the data according to the time window, and return a three-level nested list, where each sublist contains the user, project, and tag information within a time window.

[0242] Step b8: Predict user willingness to pay.

[0243] The target user representation is input into the decoder, and the neural network is regarded as a polynomial likelihood function. The parameters of the neural network can be adjusted by maximizing the likelihood estimation. The target user's preference feature representation is mapped to the target user's preference probability distribution for all flights through a polynomial decoder through a nonlinear mapping, as shown below:

[0244] π(emb u )=Softmax(σ(W d emb u +B d ))

[0245] Among them, W d is the trainable projection matrix, Bd is the trainable bias matrix.

[0246] Further calculate the target user's preference probability distribution Q for all flights, and minimize π(emb u ) and Q, and take it as the optimization target, as shown in the formula:

[0247]

[0248] Among them, M is the number of target user tables, g is the number of flights, For target user u j The preference probability for flight k.

[0249] The prediction method of user willingness to pay provided by the embodiment of the present invention introduces a characteristic evolution mechanism of the target user, so that the flight information interacting with the target user can be retained, the obtained preference characteristics of the target user are more comprehensive, and the final prediction result is more accurate, and the prediction result is more in line with the actual situation of the target user. In addition, the prediction method provided by the embodiment of the present invention also takes into account the characteristic differences of users at multiple time levels, and can construct the time series characteristics of the target user from multiple angles.

[0250] In the above-mentioned embodiment, a method for predicting a user's willingness to pay is provided. Correspondingly, the present invention also provides a device for predicting a user's willingness to pay. The device for predicting a user's willingness to pay provided in the embodiment of the present invention can implement the above-mentioned method for predicting a user's willingness to pay. The device for predicting a user's willingness to pay can be implemented by software, hardware, or a combination of software and hardware. For example, the device for predicting a user's willingness to pay can include integrated or separate functional modules or units to perform the corresponding steps in the above-mentioned methods.

[0251] Please refer to Figure 5 , which shows a schematic diagram of a user willingness to pay prediction device provided by some embodiments of the present invention. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described below is only illustrative.

[0252] like Figure 5 As shown, the user willingness to pay prediction device 500 may include:

[0253] The first generating module 501 is used to generate feature representations of multiple flight items corresponding to the flight feature layer based on the continuous numerical features and discrete category features in the flight feature information;

[0254] A second generating module 502, configured to generate corresponding session representations based on feature representations of multiple flight items;

[0255] An evolution module 503 is used to evolve the preferences of the target user at different time levels based on the session representation and feature representation evolution network to obtain the preference feature representation of the target user;

[0256] A determination module 504, for determining the probability of the target user's willingness to pay for the target flight among all flights based on the target user's preference feature representation and probability distribution;

[0257] The prediction module 505 is used to predict the target user's willingness to pay for the target flight based on the willingness to pay probability of the target flight to obtain a prediction result.

[0258] In some implementations of the embodiments of the present invention, the determination module 504 is specifically used to:

[0259] Through the polynomial decoder, the target user's preference feature representation is mapped into the target user's willingness to pay probability distribution for all flights;

[0260] Obtain a probability distribution, and use the probability distribution as a target distribution to approximate the probability distribution of willingness to pay;

[0261] Through the KL divergence method, the willingness to pay probability distribution and probability distribution are minimized and optimized to obtain the willingness to pay probability of the target flight.

[0262] In some implementations of the embodiments of the present invention, the device 500 for predicting user willingness to pay provided by the embodiments of the present invention may further include:

[0263] Probability distribution determination module (in Figure 5 ), which is used to convert the target flight into a flight number vector for the target flight in a preset dimension by using a one-hot encoding method before obtaining the probability distribution; and use the flight number vector as the probability distribution.

[0264] In some implementations of the embodiments of the present invention, the evolution module 503 is specifically used to:

[0265] Get the session representation, which is used to represent the interaction process between all flights and the target user in the session;

[0266] For the target user, the update function of the feature representation evolution network is used to express the target user and the kth session The feature evolution process when interaction occurs is to obtain the feature representation generated by the kth interaction and the preference representation generated by the k-1th interaction, where u jFor the target user, the feature evolution process is used to aggregate the long-term features and short-term change features of the target user. The neural networks used to aggregate the long-term features and short-term change features include: RNN, LSTM and GRU.

[0267] In some implementations of the embodiments of the present invention, the second generating module 502 is specifically used to:

[0268] Get flight feature sequence;

[0269] Arrange the flight feature sequences in order according to the chronological order of the time-stamped interaction events;

[0270] Obtain feature representations of multiple flight items;

[0271] Through the preset aggregation function, the feature representations of multiple flight items are mapped to corresponding session representations, and the neural network used by the preset aggregation function is LSTM.

[0272] In some implementations of the embodiments of the present invention, the first generating module 501 is specifically used for:

[0273] Acquire flight feature information, which includes information corresponding to continuous numerical features and information corresponding to discrete category features. Continuous numerical features are dense features, and discrete category features are sparse features.

[0274] For continuous numerical features, the numerical features are projected and converted into a first feature representation through a projection matrix. The first feature representation includes a projection matrix and a bias matrix.

[0275] For discrete category features, the discrete category features are numerically processed to convert the category identifiers into numbers;

[0276] The discrete category features that have been numerically processed are converted to obtain dense low-dimensional features; the second feature representation is obtained by using numbers as indexes and looking up tables;

[0277] The first feature representation and the second feature representation are obtained, and the first feature representation and the second feature representation are concatenated to generate feature representations of multiple flight items corresponding to the flight feature layer.

[0278] In some implementations of the embodiments of the present invention, the device 500 for predicting user willingness to pay provided by the embodiments of the present invention may further include:

[0279] Session partitioning module (in Figure 5 ), specifically used for:

[0280] Before generating feature representations of a plurality of flight items corresponding to the flight feature layer based on the continuous numerical features and the discrete category features in the flight feature information, an event sequence is obtained and the event sequence is divided into a session set;

[0281] For the session set, obtain the number of sessions retained by the target user; and obtain the length of events retained in a session;

[0282] Obtain a time window T, and use the time window T as a session division mark;

[0283] When the time interval between two adjacent events in the event sequence is greater than the time window T, the two adjacent events are divided into different sessions to divide the session of the target user;

[0284] The kth session S of the target user k The time when the first interaction event in the set occurs is taken as the session S k Time mark;

[0285] Set the time stamp set of the session, which includes time information at the hour level, week level, and month level.

[0286] In some implementations of the embodiments of the present invention, in some implementations of the embodiments of the present invention, the user willingness to pay prediction device 500 provided by the embodiments of the present invention is based on the same inventive concept as the user willingness to pay prediction method provided by the aforementioned embodiments of the present invention and has the same beneficial effects.

[0287] According to another embodiment, there is also provided a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute a combination of Figure 1 The method described.

[0288] According to another embodiment of the present invention, there is also provided an electronic device, comprising a memory and a processor, wherein the memory stores an executable code, and when the processor executes the executable code, the Figure 1 The method described.

[0289] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the present invention can be implemented by hardware, software, firmware or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0290] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting user willingness to pay, characterized in that: The method comprises: Based on the continuous numerical features and discrete categorical features in the flight feature information, generate feature representations of multiple flight items corresponding to the flight feature layer; generating corresponding conversation representations based on the feature representations of the plurality of flight items; Based on the session representation and feature representation evolution network, the preferences of the target user at different time levels are evolved to obtain the preference feature representation of the target user; Determining the target user's willingness to pay probability for the target flight among all flights based on the target user's preference characteristic representation and probability distribution; Based on the willingness-to-pay probability of the target flight, the willingness-to-pay of the target user for the target flight is predicted to obtain a prediction result.

2. The prediction method according to claim 1, characterized in that: The step of determining the probability of the target user's willingness to pay for the target flight among all the flights based on the target user's preference feature representation and probability distribution includes: By means of a polynomial decoder, the preference characteristic representation of the target user is mapped into a probability distribution of the target user's willingness to pay for all flights; Obtaining a probability distribution, and using the probability distribution as a target distribution to which the willingness-to-pay probability distribution is approximated; The willingness to pay probability distribution and the probability distribution are minimized and optimized by the KL divergence method to obtain the willingness to pay probability of the target flight.

3. The prediction method according to claim 2, characterized in that: Before obtaining the probability distribution, the method further includes: The target flight is converted into a flight number vector for the target flight in a preset dimension by using a one-hot encoding method; and the flight number vector is used as the probability distribution.

4. The prediction method according to claim 1, characterized in that: The evolving network based on the session representation and feature representation evolves preferences of target users at different time levels, including: Acquire the session representation, where the session representation is used to replace the interaction process between all flights in the session and the target user; For the target user, the update function of the feature representation evolution network is used to express the target user and the kth session The feature evolution process when the interaction occurs, to obtain the feature representation generated by the kth interaction and the preference representation generated by the k-1th interaction, where the kth session u j For the target user, the long-term features and short-term change features of the target user are aggregated through the feature evolution process, and the neural networks used to aggregate the long-term features and the short-term change features include: RNN, LSTM and GRU.

5. The prediction method according to claim 1, characterized in that: The generating corresponding conversation representations based on the feature representations of the plurality of flight items includes: Get flight feature sequence; Arranging the flight feature sequences in sequence according to the chronological order of the time-stamped interaction events; Obtaining feature representations of the plurality of flight items; The feature representations of the plurality of flight items are mapped to corresponding session representations through a preset aggregation function, and the neural network adopted by the preset aggregation function is LSTM.

6. The prediction method according to claim 1, characterized in that: The step of generating feature representations of a plurality of flight items corresponding to the flight feature layer based on the continuous numerical features and the discrete category features in the flight feature information includes: Acquire flight feature information, the flight feature information including information corresponding to the continuous numerical feature and information corresponding to the discrete category feature, the continuous numerical feature is a dense feature, and the discrete category feature is a sparse feature; For continuous numerical features, the numerical features are projected and converted into a first feature representation through a projection matrix, where the first feature representation includes a projection matrix and a bias matrix; For discrete category features, numerical processing is performed on the discrete category features to convert the category identifier into a number; The discrete category features processed by numerical processing are converted to obtain dense low-dimensional features; the numbers are used as indexes to obtain the second feature representation by looking up the table; The first feature representation and the second feature representation are obtained, and the first feature representation and the second feature representation are concatenated to generate feature representations of a plurality of flight items corresponding to the flight feature layer.

7. The prediction method according to claim 1, characterized in that: Before generating feature representations of a plurality of flight items corresponding to the flight feature layer based on the continuous numerical features and the discrete category features in the flight feature information, the method further includes: Acquire an event sequence, and divide the event sequence into the session set; For the session set, obtaining the number of sessions retained by the target user; and obtaining the length of events retained in one session; Obtain a time window T, and use the time window T as a session division mark; When the time interval between two adjacent events in the event sequence is greater than the time window T, the two adjacent events are divided into different sessions to divide the session of the target user; The kth session S of the target user k The time when the first interaction event in the set occurs is taken as the session S k Time mark; A time stamp set of the session is set, wherein the time stamp set includes time information at the hour level, time information at the week level, and time information at the month level.

8. A device for predicting user willingness to pay, characterized in that: The device comprises: A first generation module, used to generate feature representations of multiple flight items corresponding to the flight feature layer based on continuous numerical features and discrete category features in the flight feature information; A second generating module, configured to generate corresponding session representations based on the feature representations of the plurality of flight items; An evolution module, used to evolve the preferences of the target user at different time levels based on the session representation and feature representation evolution network to obtain the preference feature representation of the target user; A determination module, configured to determine the probability of the target user's willingness to pay for a target flight among all flights based on the target user's preference feature representation and probability distribution; The prediction module is used to predict the payment willingness of the target user for the target flight based on the payment willingness probability of the target flight to obtain a prediction result.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores executable codes, and when the processor executes the executable codes, the method according to any one of claims 1 to 7 is implemented.

Images