Dynamic flight information pushing method and system for machine learning

Abstract

The invention provides a dynamic flight information pushing method and system based on machine learning, and belongs to the technical field of artificial intelligence. The method comprises the following steps: collecting flight demand information input by a user, and forming three types of structured arrays in combination with predefined supply and demand characteristics, time sequence characteristics and temporary characteristics; inputting the supply-demand feature array and the time sequence feature array into a prediction model of machine learning training to obtain an initial price fluctuation curve; retrieving price fluctuation vector mode values of the historical flight sample sets meeting the temporary features, and setting the price fluctuation vector mode values as correction vectors; and dynamically adjusting the initial curve according to the correction vector to obtain an updated price fluctuation curve, and pushing the updated price fluctuation curve to the user according to a predicted price sequence. Through accurate prediction of the flight price, more accurate flight information pushing is provided, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a machine learning method and system for pushing dynamic flight information. Background Art

[0002] With the rapid development of the aviation industry and the widespread adoption of internet technology, flight information and price prediction services have become crucial for passengers' travel decisions. Existing flight price predictions primarily rely on historical price averages combined with fixed rules, such as discounts for purchasing tickets 21 days in advance.

[0003] However, traditional flight price forecasts rely on historical data analysis and pre-set price fluctuation rules, failing to fully account for real-time changes in market supply and demand. For example, during major holidays, emergencies, or peak seasonal travel periods, flight prices often fluctuate dramatically. Traditional flight price forecasts, lacking a real-time response mechanism to these temporary factors, struggle to accurately reflect actual price trends. Due to their inability to dynamically respond to real-time market supply and demand changes, the price information pushed by traditional flight price forecasts deviates significantly from actual price trends, impacting users' travel decisions and overall travel experience. Summary of the Invention

[0004] The present invention aims to solve the technical problem that flight price prediction in the existing technology cannot dynamically respond to real-time market supply and demand changes, resulting in inaccurate flight price information push and affecting user experience. A machine learning dynamic flight information push method and system are provided to solve the problem.

[0005] The technical solution of the present invention to solve the above technical problems is as follows:

[0006] In a first aspect, the present invention provides a dynamic flight information push method based on machine learning, comprising: based on the flight starting point, flight end point and demand time window input by the user terminal, in combination with predefined supply and demand feature attributes, timing feature attributes and temporary feature attributes, collecting supply and demand feature structured arrays, timing feature structured arrays and temporary feature structured arrays; inputting the supply and demand feature structured arrays and the timing feature structured arrays into a flight price curve prediction model trained by machine learning to obtain an initial flight price fluctuation curve; retrieving the mode value of the flight price fluctuation vector of the historical flight sample set that meets the temporary feature structured array, and setting it as a fluctuation curve correction vector; correcting the initial flight price fluctuation curve according to the fluctuation curve correction vector to obtain an updated flight price fluctuation curve, sorting the flight information from high to low according to the predicted price, and pushing it to the user terminal.

[0007] In a second aspect, the present invention provides a machine learning dynamic flight information push system, comprising: a data acquisition module for collecting supply and demand feature structured arrays, time series feature structured arrays and temporary feature structured arrays based on the flight starting point, flight end point and demand time window input by the user terminal, in combination with predefined supply and demand feature attributes, time series feature attributes and temporary feature attributes; a price prediction module for inputting the supply and demand feature structured arrays and the time series feature structured arrays into a flight price curve prediction model trained by machine learning to obtain an initial flight price fluctuation curve; a fluctuation correction module for retrieving the mode value of the flight price fluctuation vector of the historical flight sample set that meets the temporary feature structured array, and setting it as the fluctuation curve correction vector; a sorting and pushing module for correcting the initial flight price fluctuation curve according to the fluctuation curve correction vector to obtain an updated flight price fluctuation curve, sorting the flight information from high to low according to the predicted price, and pushing it to the user terminal.

[0008] The beneficial effects of the present invention are:

[0009] Based on the flight origin, destination, and demand time window input by the user, combined with predefined supply and demand, time series, and temporary feature attributes, a supply and demand feature structured array, a time series feature structured array, and a temporary feature structured array are collected. This ensures that the subsequent forecasting model fully considers various factors affecting flight prices, laying the data foundation for real-time response to market supply and demand changes. The supply and demand feature structured array and the time series feature structured array are input into a flight price curve forecasting model trained through machine learning to obtain an initial flight price fluctuation curve. This improves forecasting flexibility and accuracy, overcoming the limitations of traditional fixed-rule forecasting methods. The mode value of the flight price fluctuation vector in the historical flight sample set that satisfies the temporary feature structured array is retrieved and set as the fluctuation curve correction vector, enhancing the real-time responsiveness of the forecast. The initial flight price fluctuation curve is corrected based on the fluctuation curve correction vector to obtain an updated flight price fluctuation curve. Flight information is sorted from high to low by predicted price and pushed to the user, allowing users to quickly identify the flight options that best meet their needs, significantly improving the accuracy of flight price information.

[0010] Through the above technical solution, this application realizes dynamic and accurate prediction of flight prices, solves the problem that traditional technology cannot respond to market supply and demand changes in real time, provides users with more accurate flight price information push services, and improves user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 A flow chart of a machine learning-based dynamic flight information push method provided by the present invention;

[0012] Figure 2This is a structural diagram of a machine learning dynamic flight information push system provided by the present invention.

[0013] In the accompanying drawings, the components represented by the reference numerals are as follows:

[0014] Data collection module 11, price prediction module 12, fluctuation correction module 13, sorting and pushing module 14. DETAILED DESCRIPTION

[0015] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0016] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the specified features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0017] In the description of the present invention, the term "for example" is used to mean "used as an example, illustration or illustration". Any embodiment of the present invention described as "for example" is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed herein.

[0018] Example 1, as Figure 1 As shown, an embodiment of the present invention provides a method for pushing dynamic flight information based on machine learning, including:

[0019] S1. Based on the flight origin, flight destination, and demand time window input by the user, combined with predefined supply and demand feature attributes, timing feature attributes, and temporary feature attributes, the supply and demand feature structured array, the timing feature structured array, and the temporary feature structured array are collected.

[0020] Specifically, the system first receives travel parameters entered by the user through the user terminal, including the flight origin (i.e., departure airport), flight destination (i.e., destination airport), and desired time window (i.e., the time range within which the user wishes to travel). Based on these travel parameters, the system then collects relevant information from multiple data sources and structures it, combining it with predefined supply and demand, time series, and temporary feature attributes, to form structured arrays of supply and demand, time series, and temporary features.

[0021] Among them, the supply and demand feature structured array contains indicators reflecting the market supply and demand relationship, such as flight load factor, route competition index, and prices of alternative transportation modes; the time series feature structured array contains data elements related to time series, such as ticket purchase cycle patterns and seasonal indexes; the temporary feature structured array contains factors that may have a short-term impact on flight prices, such as fuel price futures data, extreme weather warnings, and specific types of activity data.

[0022] The structured arrays of supply and demand features, time series features, and temporary features are all organized according to a predefined data structure format to ensure that subsequent machine learning models can efficiently process this input data. By collecting data from multiple sources, a comprehensive data foundation is laid for subsequent flight price predictions.

[0023] S2. Input the supply and demand feature structured array and the time series feature structured array into a flight price curve prediction model trained by machine learning to obtain an initial flight price fluctuation curve.

[0024] Specifically, the structured arrays of supply and demand characteristics and time series characteristics are fed as input parameters into a flight price curve prediction model pre-trained using machine learning methods. After receiving these two types of structured data, the flight price curve prediction model performs forward calculations and inference based on its internally trained parameter weights and network structure, ultimately outputting an initial flight price fluctuation curve.

[0025] The flight price curve prediction model is a composite model that integrates supply and demand characteristics with time series features. Its internal structure comprises a base prediction model based on supply and demand characteristics, a base prediction model based on time series characteristics, and a meta-model that fuses the outputs of these two. By learning from a large amount of historical flight price data, this model understands the inherent patterns in flight price fluctuations as a function of supply and demand and time series.

[0026] The flight price curve prediction model analyzes factors such as flight load factor, route competition index, and prices of alternative transportation modes in the supply and demand feature structured array. At the same time, it combines time dimension information such as ticket purchase cycle pattern and seasonal index in the time series feature structured array to comprehensively evaluate the degree and direction of the impact of these factors on flight prices.

[0027] The flight price curve prediction model calculates and processes the initial flight price fluctuation curve, which describes the expected change trend of flight prices within the demand time window. This initial flight price fluctuation curve includes price forecasts based on supply and demand relationships and time series patterns, and has stronger dynamic response capabilities than traditional forecasting methods that rely on historical averages.

[0028] S3. Retrieve the mode value of the flight price fluctuation vector of the historical flight sample set that meets the temporary feature structured array, and set it as the fluctuation curve correction vector.

[0029] Specifically, a similarity match is first performed within the historical flight database, using the obtained temporary feature structured array as a benchmark. When the similarity between the temporary feature structured array of a historical sample in the historical flight database and the temporary feature structured array of the current query is greater than or equal to a preset similarity threshold, the historical sample is added to the historical flight sample set. This similarity is calculated using standard metrics such as cosine similarity or Euclidean distance to ensure that the selected historical sample has a high degree of similarity with the current query criteria in terms of the temporary feature dimension.

[0030] Secondly, for the selected historical flight sample set, the flight price fluctuation vector of each historical sample is extracted to form a set of flight price fluctuation vectors. Each fluctuation vector in this set describes the degree and direction of deviation of the flight price from the benchmark price under specific temporary characteristic conditions. Subsequently, a box plot analysis is performed on the set of flight price fluctuation vectors to identify and eliminate outliers, and obtain the flight price fluctuation vectors within the box. The box plot analysis method can effectively filter out extreme abnormal samples and improve the stability and reliability of fluctuation vector statistics. Afterwards, the mean of the flight price fluctuation vectors within the box is calculated and set as the fluctuation curve correction vector. This fluctuation curve correction vector represents the average deviation of the flight price from the conventional expectation under the current temporary characteristic conditions and will be used to correct the initial flight price fluctuation curve.

[0031] By obtaining the fluctuation curve correction vector, flight price fluctuation patterns similar to current temporary features can be extracted from historical data. These patterns reflect the impact of temporary factors such as fuel price changes, extreme weather and special events on flight prices, thereby effectively utilizing the correlation between temporary features and price fluctuations in historical data, providing a data-driven correction basis for subsequent price forecasts.

[0032] S4. Correct the initial flight price fluctuation curve according to the fluctuation curve correction vector to obtain an updated flight price fluctuation curve, sort the flight information from high to low according to the predicted price, and push the updated flight price fluctuation curve to the user end.

[0033] Specifically, the volatility curve correction vector is first applied to the initial flight price volatility curve, performing a vector addition operation. During this operation, each component of the volatility curve correction vector is applied to the predicted value of the initial flight price volatility curve at the corresponding time point, thereby making a fine point-by-point correction to the curve and obtaining an updated flight price volatility curve. The initial flight price volatility curve reflects the price trend based on supply and demand characteristics and time series characteristics, while the volatility curve correction vector supplements the additional impact of temporary factors such as fuel price fluctuations, extreme weather, and special events. This vector superposition method achieves a comprehensive consideration of multi-dimensional features, allowing the final forecast to more comprehensively reflect the combined impact of various factors on flight prices.

[0034] Subsequently, based on the updated flight price fluctuation curve, the predicted price of each flight within the user's demand time window is calculated. All feasible flights that meet the user's query conditions (i.e., flight origin, flight destination, and demand time window) are sorted from high to low according to the predicted price. The sorting process takes into account the price prediction values ​​of different flights, so that users can intuitively understand the price competitiveness of each flight. Afterwards, the sorted flight information is pushed to the user end for display. The push content includes basic flight information (such as flight number, departure time, arrival time, airline, etc.) and the corresponding predicted price information. The user-end interface presents this information in a clear manner, and may provide price trend charts to help users intuitively understand price change trends, providing users with dynamic and accurate flight price prediction services, so that they can make more economical ticket purchase decisions. Compared with traditional prediction methods, this embodiment can respond to market supply and demand changes and temporary special factors in real time, significantly improving prediction accuracy and user experience.

[0035] Furthermore, the flight price curve prediction model is constructed by the following steps:

[0036] S21. Based on the supply and demand characteristic attributes, construct a supply and demand characteristic flight price prediction base model through machine learning;

[0037] S22. Based on the time series feature attributes, construct a time series feature flight price prediction base model through machine learning;

[0038] S23. Using the outputs of the supply-demand feature flight price prediction base model and the time series feature flight price prediction base model as input, construct a flight price prediction meta-model through machine learning;

[0039] S24. Merge the output layer of the base model with the input layer of the meta-model to obtain the flight price curve prediction model.

[0040] In one feasible implementation, a specialized base model for flight price prediction based on supply and demand characteristics is first constructed using machine learning methods, utilizing supply and demand characteristic attributes, including core ones (such as flight load factor, route competition index, and alternative transportation mode prices) and selected marginal ones. This base model utilizes a multi-branch neural network structure, in which each branch is responsible for processing a specific supply and demand characteristic. Ultimately, the feature representations of each branch are integrated through a fully connected layer to establish a mapping relationship between supply and demand characteristics and flight prices. Transfer learning techniques are applied during the training of the base model, first loading a pre-trained generalized model and then fine-tuning it for local data, thereby improving model deployment efficiency and prediction accuracy. Simultaneously, a specialized base model for flight price prediction based on time series characteristics is constructed using machine learning methods, utilizing core ones (such as ticket purchase cycle patterns and seasonality indices) and selected marginal ones. This base model can employ a time series modeling approach combined with an LSTM (Long Short-Term Memory) or Transformer architecture to capture the long- and short-term dependencies of flight prices over time. Similar to the supply-demand feature flight price prediction base model, the time series feature flight price prediction base model also uses transfer learning technology to achieve rapid deployment and efficient adaptation by preloading the generalized model and performing local fine-tuning.

[0041] Next, a flight price prediction meta-model is constructed to integrate prediction results from both supply and demand characteristics and time series characteristics. Specifically, the outputs of the supply and demand and time series flight price prediction base models are used as input features, and a machine learning approach is used to construct the flight price prediction meta-model. This flight price prediction meta-model learns the optimal combined weights of the outputs of the two base models. It dynamically adjusts the importance of supply and demand and time series factors in the final prediction based on the characteristics of different routes and time periods. The flight price prediction meta-model can be trained using a neural network structure or ensemble learning method by minimizing prediction error to obtain the optimal feature fusion strategy. Subsequently, the output layers of the supply and demand and time series flight price prediction base models are merged and connected with the input layer of the flight price prediction meta-model to construct a complete flight price curve prediction model. This merging process ensures lossless information flow, allowing the representations of supply and demand and time series characteristics to be fully integrated and complementary. The resulting flight price curve prediction model simultaneously considers market supply and demand factors and time series factors, producing a more comprehensive and accurate forecast of flight price fluctuations.

[0042] Through this multi-stage construction process, a highly flexible and adaptable flight price prediction model architecture was achieved. In particular, the use of transfer learning technology enabled the rapid construction of localized models adapted to specific regions or routes based on a universal pre-trained model. This significantly reduced model deployment time and computing resource requirements, enabling rapid local deployment while maintaining high-precision predictions, meeting the practical application needs of diverse market environments.

[0043] Furthermore, the predefined supply and demand characteristic attributes, time series characteristic attributes and temporary characteristic attributes are combined, including:

[0044] S11. Configuring core supply and demand characteristic attributes and marginal supply and demand characteristic attributes, and setting them as the supply and demand characteristic attributes, wherein the core supply and demand characteristic attributes include flight load factor, route competition index, and alternative transportation mode price, and the marginal supply and demand characteristic attributes are user-defined.

[0045] S12. Configure core time series feature attributes and edge time series feature attributes, and set them as the time series feature attributes, wherein the core time series feature attributes include a ticket purchase cycle pattern and a seasonal index, and the edge time series feature attributes are user-defined.

[0046] S13. Configure core temporary feature attributes and edge temporary feature attributes, and set them as the temporary feature attributes. The core temporary feature attributes include fuel price futures data, extreme weather warnings, and set type activity data. The edge temporary feature attributes are customized by the user end.

[0047] In a preferred embodiment, supply and demand attributes are categorized into core and marginal supply and demand attributes. Core supply and demand attributes are widely validated key supply and demand indicators with a significant and universal impact on flight prices. They include flight load factor, route competition index, and alternative transportation prices. The flight load factor reflects seat occupancy on a specific flight and is calculated by collecting real-time data on remaining seats. This metric reflects the balance between supply and demand on a flight. A high flight load factor indicates tight seat supply, which results in a corresponding increase in price forecasts. Conversely, a low flight load factor indicates oversupply, which results in a decrease in price forecasts. The route competition index quantifies the number of airlines operating on the same route and their market share. It is calculated using a route database and market share analysis. This metric reflects the level of competition on a route. A higher competition index indicates a greater likelihood of price suppression due to competition, which results in corresponding adjustments to price forecasting strategies. Alternative transportation prices include dynamic price data for alternative modes of travel, such as high-speed rail and long-distance buses, between the same departure and destination points. This metric is collected via a real-time API or periodic crawlers and is used to assess price competition between different transportation modes, providing a market reference for flight pricing. Edge supply and demand characteristic attributes allow for supplementary supply and demand indicators that can be customized through user-side configuration, such as the economic index of a specific region, the destination tourism popularity index, etc. By providing a standardized interface, users can flexibly expand the supply and demand characteristic dimensions according to different market environments.

[0048] Time series feature attributes are categorized into core and marginal time series attributes. Core time series attributes are key time series indicators proven to have significant explanatory power for temporal flight price fluctuations, including the ticket purchase cycle pattern and seasonality index. The ticket purchase cycle pattern describes the changing patterns in booking acceleration during different pre-departure time periods and is derived by analyzing the temporal distribution characteristics of historical booking data. This metric can capture purchasing behavior patterns at different lead times, such as booking surges at key time points like 7, 14, and 21 days before departure, providing an important basis for predicting temporal flight price fluctuations. The seasonality index quantifies the multiplier effect of special periods such as holidays, winter and summer vacations on flight prices and is calculated through seasonal decomposition of historical price data. This metric reflects demand differences across time periods, such as the price multiplier effect during Spring Festival and Golden Week, and can be used to adjust price forecasts for corresponding periods accordingly. Marginal time series attributes are supplementary time series indicators that can be customized by the user, such as weekday / weekend patterns and cyclical changes in the business / leisure travel ratio. By providing a standardized interface, users can flexibly extend the time series feature dimensions based on different route characteristics.

[0049] Temporary attributes are categorized into core and marginal temporary attributes. Core attributes are key temporary factors proven to have a significant short-term impact on flight prices, including fuel price futures data, extreme weather warnings, and event data. Fuel price futures data refers to fuel-related futures trading data, acquired in real time through a financial data API. This metric is directly linked to airline operating costs. Rising fuel prices typically lead to short-term increases in flight prices. Monitoring the fuel futures market can be used to predict cost-driven changes in flight prices. Extreme weather warnings access a meteorological API to obtain extreme weather warning information for departure, destination, and route locations. This metric is used to predict weather events that may cause flight delays, cancellations, or sudden changes in demand, allowing price fluctuation expectations to be adjusted based on the severity of the weather anomaly. Event data uses a specially designed crawler system to collect information on major events at the destination, such as concerts, exhibitions, and sporting events. This metric is used to identify events that may lead to short-term demand surges on specific routes, allowing for the prediction of temporary price increases. Edge temporary feature attributes allow for supplementary temporary indicators that can be customized through user-side configuration, such as exchange rate fluctuations, public emergencies, etc. By providing standardized interfaces, users can flexibly expand temporary feature dimensions according to different market environments.

[0050] Through the above-mentioned feature attribute configuration, a flexible and scalable multi-dimensional feature system was constructed. This system divides the factors affecting flight prices into three dimensions: supply and demand relationship, time pattern and temporary factors. It also distinguishes between core features and marginal features within each dimension, providing a basis for achieving accurate dynamic predictions. This is superior to traditional prediction methods that only rely on historical prices and fixed rules.

[0051] Furthermore, based on the supply and demand characteristic attributes, a supply and demand characteristic flight price prediction base model is constructed through machine learning, including:

[0052] S211. Downloading a generalized model for flight price prediction based on supply and demand characteristics constructed by machine learning based on the core supply and demand characteristic attributes from a preset model library;

[0053] S212. Sorting the marginal supply and demand characteristic attributes for flight price relevance to obtain selected marginal supply and demand characteristic attributes;

[0054] S213. Based on the selected marginal supply and demand feature attributes and the core supply and demand feature attributes, transfer learning is performed on the supply and demand feature flight price prediction generalization model to obtain the supply and demand feature flight price prediction base model.

[0055] In a preferred embodiment, first, a pre-trained model reuse technology is used to download a generalized model for flight price prediction based on supply and demand characteristics that has been trained with large-scale data from a preset model library. The generalized model for flight price prediction based on supply and demand characteristics is a general prediction model constructed through deep learning methods based on core supply and demand characteristic attributes (i.e., flight load factor, route competition index, and price of alternative transportation modes), and has been trained and verified on a large amount of route data. The preset model library stores a variety of pre-trained model variants for different regional characteristics and different route types. According to the characteristics of the current application scenario, the most matching generalized model is selected for download. This pre-trained model reuse mechanism reduces the computing resource requirements for local model training and accelerates the model deployment process. The downloaded generalized model for flight price prediction based on supply and demand characteristics has learned the basic mapping relationship between core supply and demand characteristics and flight prices, providing good parameter initialization for subsequent localized adjustments.

[0056] Then, the flight price correlation analysis and screening are performed on the edge supply and demand feature attributes configured by the user to determine which edge supply and demand features should be included in the model training process. Specifically, first, local historical flight data within a certain time window is collected, including flight prices and the values ​​of all edge supply and demand features; then, the Pearson correlation coefficient, Spearman rank correlation coefficient or mutual information value between each edge supply and demand feature and the flight price is calculated to quantify the degree of linear or nonlinear correlation between the feature and the target variable and determine the correlation index; based on the calculated correlation index, a correlation threshold is set (which can be dynamically adjusted according to the application scenario), and edge supply and demand features whose correlation exceeds the threshold are screened out and marked as selected edge supply and demand feature attributes. Through the correlation sorting step, the edge supply and demand features that are most valuable for local flight price prediction can be identified and retained, and irrelevant or redundant features can be filtered out, thereby improving model training efficiency and prediction accuracy.

[0057] Subsequently, transfer learning is performed on the downloaded generalized supply and demand feature flight price prediction model based on the selected marginal supply and demand feature attributes and predefined core supply and demand feature attributes, constructing a base model for supply and demand feature flight price prediction that is adapted to local market characteristics. The transfer learning process consists of two stages: First, based on the number and types of selected marginal supply and demand feature attributes, the input layer of the generalized supply and demand feature flight price prediction model is expanded to construct additional feature processing branches. In this stage, the original parameters of the generalized supply and demand feature flight price prediction model are frozen, and only the newly added feature processing branches are trained, enabling the model to effectively handle the local marginal supply and demand features. Second, some parameters of the generalized supply and demand feature flight price prediction model are unfrozen, and the entire model is fine-tuned with a low learning rate to adapt the core feature processing logic to local market characteristics. During the transfer learning process, locally collected flight price data is used as the training set, and batch gradient descent or its variants are employed to minimize the mean squared error or mean absolute percentage error between the predicted and actual prices, thereby optimizing the model parameters and obtaining the base model for supply and demand feature flight price prediction. Through transfer learning, it is possible to quickly adapt to the particularities of the local market while retaining the general knowledge of the generalized model, significantly improving prediction accuracy.

[0058] The above steps significantly improve model deployment efficiency, reduce computing resource requirements, and ensure prediction accuracy. Furthermore, the methods used in this embodiment to construct the base model for flight price prediction based on time series features and the base model for flight price prediction based on temporary features are similar to the above process, both employing the basic strategy of downloading a pre-trained model, sorting feature relevance, and performing transfer learning. However, the specific implementation may be adjusted based on the different feature types.

[0059] Furthermore, the steps of constructing the generalized model for flight price prediction based on supply and demand characteristics include:

[0060] S2111. Based on the number of core supply and demand characteristic attributes, construct an equal number of fully connected branch neural networks;

[0061] S2112, build a fully connected neural network;

[0062] S2113. Using the output of the fully connected branch neural network as the input of the fully connected neural network, and constructing a generalized model architecture for flight price prediction based on supply and demand characteristics;

[0063] S2114. Using the core supply and demand feature attributes as the only variable, collect the core supply and demand feature value dataset and the flight price identification dataset, train the supply and demand feature flight price prediction generalization model architecture, and obtain the supply and demand feature flight price prediction generalization model.

[0064] In a preferred embodiment, first, based on the predefined number of core supply and demand feature attributes, an equal number of fully connected branch neural networks are constructed. Since the core supply and demand feature attributes include three indicators: flight load factor, route competition index, and price of alternative transportation modes, three independent fully connected branch neural networks are constructed accordingly. Each branch network specializes in processing a type of core supply and demand feature and can effectively capture the nonlinear relationship between the feature and the flight price. Each branch neural network adopts a similar structure, usually containing 2-3 hidden layers, with an appropriate number of neurons (for example, 32-128) set in each layer. Different branches can appropriately adjust the network depth and width according to the complexity of the corresponding features; each hidden layer is followed by a ReLU activation function to enhance the nonlinear expression ability of the network. At the same time, batch normalization technology is applied to accelerate training convergence, and a Dropout layer is added after some hidden layers (the dropout rate is usually set to 0.2-0.5) to prevent overfitting. This multi-branch parallel processing network design enables the model to extract feature representations for different types of supply and demand features separately, effectively handle the heterogeneity of various features, and lay the foundation for subsequent feature fusion.

[0065] Then, a fully connected neural network is constructed as the feature fusion layer to integrate the feature representations extracted by each fully connected branch neural network. The fully connected neural network usually contains 2-4 hidden layers, and the number of neurons in each layer decreases layer by layer (for example, 256 neurons in the first layer and 64 neurons in the last layer), forming a funnel-shaped structure. The fusion network adopts a fully connected approach to ensure that the features of each branch can fully interact. Each hidden layer is followed by a ReLU activation function, and L2 regularization is applied (the weight decay coefficient is usually set to 0.0001-0.001) to enhance the generalization ability of the model. The last layer of the network is the output layer of a single neuron, which uses a linear activation function to directly output the predicted flight price value. This feature fusion network design enables the model to learn the interaction pattern between different supply and demand characteristics, capture the law of flight price changes under the combined influence of multiple factors, and improve prediction accuracy and interpretability.

[0066] The outputs of the fully connected branch neural networks are then used as inputs to a fully connected neural network, constructing a complete generalized model architecture for flight price prediction based on supply and demand characteristics. The model's forward propagation process is as follows: First, each core supply and demand feature attribute is input into its corresponding branch neural network. After undergoing multiple layers of nonlinear transformations, feature representation vectors are obtained. These feature representation vectors are then concatenated to form a unified high-dimensional feature vector, which serves as the input to the fusion network. The fusion network then further processes and reduces the dimensionality of the high-dimensional feature vector, outputting the final flight price prediction. During the construction process, advanced network components such as residual connections and attention mechanisms can be designed to enhance the model's expressiveness and training stability. Residual connections add skip connections within the branch networks, alleviating the vanishing gradient problem in deep networks. The attention mechanism calculates weight coefficients for different features, enabling the model to dynamically adjust the importance of each feature in different scenarios. This hierarchical model architecture ensures sufficient model complexity to capture complex nonlinear relationships while maintaining well-structured properties to enhance interpretability and generalization.

[0067] Subsequently, using core supply and demand feature attributes as unique variables, a large-scale dataset of core supply and demand feature values ​​and corresponding flight price identifiers was collected to train the constructed generalized model architecture for flight price prediction based on supply and demand features, resulting in a generalized model for flight price prediction based on supply and demand features. During the training data collection process, data was collected from multiple major aviation markets, covering different regions, airlines, and route types to ensure data diversity and representativeness. The collected data was preprocessed, including outlier detection and processing, missing value imputation, and feature normalization (Z-score or Min-Max normalization) to improve data quality. The generalized model architecture for flight price prediction based on supply and demand features was then trained on the preprocessed data. Mini-batch stochastic gradient descent or its optimized variants was used for model training, with the training objective being to minimize a loss function between predicted and actual prices. The loss function can be selected from mean squared error, mean absolute error, or mean absolute percentage error, depending on the application scenario. The fully trained generalized model for flight price prediction based on supply and demand features effectively captures the universal mapping between core supply and demand features and flight prices, laying the foundation for subsequent transfer learning and localized deployment. It is worth noting that the generalized model for flight price prediction based on supply and demand characteristics is trained only based on core supply and demand characteristics and does not contain any marginal feature information. This ensures the universality of the model and enables it to be applied as a general basic model in various market environments.

[0068] Through the above four steps, a generalized model for flight price prediction based on supply and demand characteristics with stable performance was established. This model is stored in the preset model library as a pre-trained basic model and can be downloaded and used in various application scenarios, thereby improving deployment efficiency and adaptability.

[0069] Furthermore, based on the selected marginal supply and demand feature attributes and the core supply and demand feature attributes, transfer learning is performed on the supply and demand feature flight price prediction generalization model to obtain the supply and demand feature flight price prediction base model, including:

[0070] S2131. Based on the number of selected edge supply and demand feature attributes, construct an equal number of incremental fully connected branch neural networks;

[0071] S2132: Using the output of the incremental fully connected branch neural network as the input of the fully connected neural network to construct a supply and demand characteristic flight price prediction base model architecture;

[0072] S2133. Freeze the fully connected branch neural network, use the selected edge supply and demand feature attribute as the only variable, collect the selected edge supply and demand feature value dataset and the flight price identification dataset, train the supply and demand feature flight price prediction base model architecture, and obtain a one-stage supply and demand feature flight price prediction base model;

[0073] S2134. Activate the fully connected branch neural network, use the core supply and demand feature attributes and the selected edge supply and demand feature attributes as unique variables, collect the supply and demand feature value dataset and the flight price identification dataset, train the one-stage supply and demand feature flight price prediction base model, and obtain the supply and demand feature flight price prediction base model.

[0074] In a preferred method, first, an equal number of incremental fully connected branch neural networks are constructed based on the number of selected edge supply and demand feature attributes. Each incremental fully connected branch neural network is responsible for processing a class of selected edge supply and demand features, converting the original features into high-level feature representations. These incremental fully connected branch neural networks adopt a structure similar to the branch network for processing core supply and demand features, usually containing 2-3 hidden layers, with an appropriate number of neurons (such as 16-64) configured in each layer. The ReLU activation function and Dropout technology (discarding rate 0.2-0.3) are applied after each hidden layer. This multi-branch parallel processing design enables the model to effectively process the edge supply and demand features unique to the local market, providing supplementary information for subsequent feature fusion.

[0075] Then, the output of the incremental fully connected branch neural network is combined with the output of the original supply and demand feature flight price prediction generalization model as the input of the fully connected neural network to construct the expanded supply and demand feature flight price prediction base model architecture. During this architectural expansion process, the overall structure of the original generalized model remains unchanged, and only necessary dimensional adjustments are made to the feature input layer and feature fusion layer. Specifically, the input dimension of the original fully connected neural network is expanded to accommodate the additional feature representations from the incremental fully connected branch neural network. In this way, the core feature processing logic learned in the generalized model is retained, and the ability to process local unique edge features is added, forming a model architecture that is both universal and targeted.

[0076] Subsequently, parameter freezing was used to perform the first phase of transfer learning on the expanded supply and demand feature flight price prediction base model architecture. In this phase, the fully connected branch neural network (the part that processes the core supply and demand features) and some fully connected neural network parameters in the original generalized model were frozen, and only the newly added incremental fully connected branch neural network and the extended parameters of the fully connected neural network were trained. During the training process, the selected edge supply and demand feature attributes were used as the only variables, and the local edge feature dataset and the corresponding flight price identification dataset were collected. This method allows the model to focus on learning the mapping relationship between edge features and flight prices while maintaining the original core feature processing capabilities. A small learning rate (usually 0.0001-0.001) and appropriate regularization were used for training to ensure that the model parameter adjustments were smooth and effective. Through the first phase of learning, a one-stage supply and demand feature flight price prediction base model that was initially adapted to the local area was obtained. The model has the ability to process local edge features, but the core feature processing logic has not yet been localized.

[0077] Subsequently, the previously frozen fully connected branched neural network is activated to comprehensively fine-tune the first-stage supply and demand feature flight price prediction base model, obtaining the final supply and demand feature flight price prediction base model. During this stage, both the core supply and demand feature attributes and selected marginal supply and demand feature attributes are used as variables, and a complete supply and demand feature dataset and a flight price identification dataset are collected. During training, all parameters of the first-stage supply and demand feature flight price prediction base model are jointly optimized, but a smaller learning rate (typically 0.00001-0.0001) is used for the original generalized model, while a larger learning rate (typically 0.0001-0.001) is maintained for the newly added components. This differentiated learning rate strategy ensures a balance between retaining existing knowledge and absorbing new knowledge. Through the second stage of comprehensive fine-tuning, a supply and demand feature flight price prediction base model that fully adapts to local characteristics is obtained. This supply and demand feature flight price prediction base model not only retains the general rules learned in the generalized model, but also integrates local special patterns. It can accurately handle the combined impact of core supply and demand features and marginal supply and demand features, providing high-precision basic support for flight price prediction.

[0078] Through transfer learning in the above steps, we achieve an efficient transition from a generalized model to a localized, specialized model, significantly reducing the data requirements and computing resource consumption for local model construction while ensuring the model's predictive accuracy and adaptability. Compared to traditional training from scratch, transfer learning enables the rapid deployment of high-performance predictive models even with limited data.

[0079] Furthermore, the mode value of the flight price fluctuation vector of the historical flight sample set that meets the temporary feature structured array is retrieved and set as the fluctuation curve correction vector, including:

[0080] S31. When the similarity between the temporary feature structured array of the historical sample and the temporary feature structured array is greater than or equal to the similarity threshold, add the historical sample to the historical flight sample set;

[0081] S32, performing a box plot analysis on the flight price fluctuation vector set of the historical flight sample set to obtain a box flight price fluctuation vector;

[0082] S33. Calculate the mean of the box flight price fluctuation vector and set it as the fluctuation curve correction vector.

[0083] In a preferred embodiment, first, similar samples are screened from the historical database using similarity matching. The similarity between the temporary feature structured array of the historical sample and the temporary feature structured array of the current query is calculated. The similarity calculation can use the cosine similarity method to effectively measure the directional similarity between the two arrays. At the same time, a similarity threshold is pre-set (usually 0.7-0.9, which can be dynamically adjusted according to the application scenario). When the calculated similarity is greater than or equal to the similarity threshold, the historical sample is added to the historical flight sample set. This similarity-based screening mechanism ensures that the selected samples are highly matched with the current query conditions in the temporary feature dimension, providing a reliable data basis for subsequent price fluctuation analysis.

[0084] Then, a boxplot analysis is performed on the flight price fluctuation vectors of the historical flight sample set to identify and filter outliers, resulting in a box-shaped flight price fluctuation vector. Boxplot analysis can intuitively display data distribution characteristics and identify outliers. Specifically, the quartiles of each component in the flight price fluctuation vector set are first calculated: the first quartile (25th percentile), the median (50th percentile), and the third quartile (75th percentile). The interquartile range is then calculated, and upper and lower bounds are defined. For each vector in the flight price fluctuation vector set, its components are checked to see if they fall within the corresponding bounds. If any component of a vector exceeds the bounds, it is deemed an outlier and removed from the set. The remaining vectors constitute the flight price fluctuation vectors within the box, known as the box-shaped flight price fluctuation vectors. This boxplot-based analysis mechanism effectively eliminates interference from extreme price fluctuation samples, improving the stability and reliability of the subsequent calculation of the correction vector.

[0085] Subsequently, the mean of the price fluctuation vectors for the boxed flights is calculated and set as the fluctuation curve correction vector. This calculation uses a vector mean algorithm, which sums the corresponding components of all boxed flight price fluctuation vectors and divides them by the number of vectors to obtain the average vector. Each component of the fluctuation curve correction vector corresponds to the price correction at a specific point in time within the demand time window. These corrections reflect the expected impact of current temporary characteristic conditions (such as fuel price futures data, extreme weather warnings, and data on designated activity types) on flight prices and will be used to make targeted adjustments to the initial flight price fluctuation curve.

[0086] Through the above steps, price fluctuation patterns similar to current temporary characteristic conditions can be extracted from historical data and applied to current forecasts. Compared with traditional methods, this method has stronger adaptability and accuracy, can effectively capture the complex impact of various temporary factors on flight prices, and provide an accurate correction basis for the final price forecast.

[0087] Example 2, as Figure 2As shown, based on the same inventive concept as the method for pushing dynamic flight information based on machine learning provided in Example 1, an embodiment of the present invention further provides a system for pushing dynamic flight information based on machine learning, including:

[0088] The data collection module 11 is used to collect the supply and demand feature structured array, the time series feature structured array, and the temporary feature structured array based on the flight origin, flight destination, and demand time window input by the user terminal and in combination with the predefined supply and demand feature attributes, time series feature attributes, and temporary feature attributes;

[0089] A price prediction module 12 is configured to input the supply and demand feature structured array and the time series feature structured array into a flight price curve prediction model trained by machine learning to obtain an initial flight price fluctuation curve;

[0090] A fluctuation correction module 13 is used to retrieve the mode value of the flight price fluctuation vector of the historical flight sample set that meets the temporary feature structured array and set it as the fluctuation curve correction vector;

[0091] The sorting and pushing module 14 is used to correct the initial flight price fluctuation curve according to the fluctuation curve correction vector, obtain an updated flight price fluctuation curve, sort the flight information from high to low according to the predicted price, and push it to the user end.

[0092] Furthermore, the steps of constructing the flight price curve prediction model include:

[0093] Based on the supply and demand characteristic attributes, a supply and demand characteristic flight price prediction base model is constructed through machine learning;

[0094] Based on the time series feature attributes, a time series feature flight price prediction base model is constructed through machine learning;

[0095] Taking the outputs of the supply-demand feature flight price prediction base model and the time series feature flight price prediction base model as input, constructing a flight price prediction meta-model through machine learning;

[0096] The output layer of the base model is merged with the input layer of the meta-model to obtain the flight price curve prediction model.

[0097] Furthermore, the data acquisition module 11 includes the following execution steps:

[0098] Configuring core supply and demand characteristic attributes and marginal supply and demand characteristic attributes, and setting them as the supply and demand characteristic attributes, wherein the core supply and demand characteristic attributes are flight load factor, route competition index, and alternative transportation mode price, and the marginal supply and demand characteristic attributes are custom configured by the user end;

[0099] Configuring core time series feature attributes and edge time series feature attributes, and setting them as the time series feature attributes, wherein the core time series feature attributes are a ticket purchase cycle pattern and a seasonal index, and the edge time series feature attributes are custom configured by the user end;

[0100] The core temporary feature attributes and the edge temporary feature attributes are configured and set as the temporary feature attributes, wherein the core temporary feature attributes are fuel price futures data, extreme weather warnings, and set type activity data, and the edge temporary feature attributes are customized by the user end.

[0101] Furthermore, the steps of constructing the supply and demand characteristic flight price prediction base model include:

[0102] Download a generalized model for flight price prediction based on supply and demand characteristics constructed through machine learning from the preset model library;

[0103] Sorting the marginal supply and demand feature attributes for flight price relevance to obtain selected marginal supply and demand feature attributes;

[0104] Based on the selected marginal supply and demand feature attributes and the core supply and demand feature attributes, transfer learning is performed on the supply and demand feature flight price prediction generalization model to obtain the supply and demand feature flight price prediction base model.

[0105] Furthermore, the steps of constructing the generalized model for flight price prediction based on supply and demand characteristics include:

[0106] Based on the number of core supply and demand feature attributes, an equal number of fully connected branch neural networks are constructed;

[0107] Build a fully connected neural network;

[0108] Using the output of the fully connected branch neural network as the input of the fully connected neural network to construct a generalized model architecture for flight price prediction based on supply and demand characteristics;

[0109] Taking the core supply and demand feature attributes as the only variable, the core supply and demand feature value dataset and the flight price identification dataset are collected, the supply and demand feature flight price prediction generalization model architecture is trained, and the supply and demand feature flight price prediction generalization model is obtained.

[0110] Furthermore, the steps of constructing the supply and demand characteristic flight price prediction base model also include:

[0111] Based on the number of selected edge supply and demand feature attributes, an equal number of incremental fully connected branch neural networks are constructed;

[0112] Using the output of the incremental fully connected branch neural network as the input of the fully connected neural network to construct a supply and demand characteristic flight price prediction base model architecture;

[0113] Freeze the fully connected branch neural network, use the selected edge supply and demand feature attribute as the only variable, collect the selected edge supply and demand feature value dataset and the flight price identification dataset, train the supply and demand feature flight price prediction base model architecture, and obtain a one-stage supply and demand feature flight price prediction base model;

[0114] Activate the fully connected branch neural network, use the core supply and demand feature attributes and the selected edge supply and demand feature attributes as unique variables, collect supply and demand feature value datasets and flight price identification datasets, train the one-stage supply and demand feature flight price prediction base model, and obtain the supply and demand feature flight price prediction base model.

[0115] Furthermore, the fluctuation correction module 13 includes the following execution steps:

[0116] When the similarity between the temporary feature structured array of the historical sample and the temporary feature structured array is greater than or equal to the similarity threshold, the historical sample is added to the historical flight sample set;

[0117] Performing a box plot analysis on the flight price fluctuation vector set of the historical flight sample set to obtain a box flight price fluctuation vector;

[0118] The mean of the box flight price fluctuation vector is calculated and set as the fluctuation curve correction vector.

[0119] It should be noted that, in the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0120] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0121] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0122] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0123] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0124] Although preferred embodiments of the present invention have been described, additional changes and modifications to these embodiments may occur to those skilled in the art once the basic inventive concepts become known.

[0125] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A machine learning method for pushing dynamic flight information, characterized in that: include: Based on the flight origin, flight destination, and demand time window input by the user, combined with predefined supply and demand feature attributes, time series feature attributes, and temporary feature attributes, supply and demand feature structured arrays, time series feature structured arrays, and temporary feature structured arrays are collected; Inputting the supply and demand feature structured array and the time series feature structured array into a flight price curve prediction model trained by machine learning to obtain an initial flight price fluctuation curve; Retrieving the mode value of the flight price fluctuation vector of the historical flight sample set that meets the temporary feature structured array and setting it as the fluctuation curve correction vector; The initial flight price fluctuation curve is corrected according to the fluctuation curve correction vector to obtain an updated flight price fluctuation curve, and the flight information is sorted from high to low according to the predicted price and pushed to the user end.

2. The method according to claim 1, wherein The flight price curve prediction model is constructed by the following steps: Based on the supply and demand characteristic attributes, a supply and demand characteristic flight price prediction base model is constructed through machine learning; Based on the time series feature attributes, a time series feature flight price prediction base model is constructed through machine learning; Taking the outputs of the supply-demand feature flight price prediction base model and the time series feature flight price prediction base model as input, constructing a flight price prediction meta-model through machine learning; The output layer of the base model is merged with the input layer of the meta-model to obtain the flight price curve prediction model.

3. The method according to claim 2, wherein Combined with predefined supply and demand characteristic attributes, timing characteristic attributes, and temporary characteristic attributes, including: Configuring core supply and demand characteristic attributes and marginal supply and demand characteristic attributes, and setting them as the supply and demand characteristic attributes, wherein the core supply and demand characteristic attributes are flight load factor, route competition index, and alternative transportation mode price, and the marginal supply and demand characteristic attributes are custom configured by the user end; Configuring core time series feature attributes and edge time series feature attributes, and setting them as the time series feature attributes, wherein the core time series feature attributes are a ticket purchase cycle pattern and a seasonal index, and the edge time series feature attributes are custom configured by the user end; The core temporary feature attributes and the edge temporary feature attributes are configured and set as the temporary feature attributes, wherein the core temporary feature attributes are fuel price futures data, extreme weather warnings, and set type activity data, and the edge temporary feature attributes are customized by the user end.

4. The method according to claim 3, wherein Based on the supply and demand characteristic attributes, a supply and demand characteristic flight price prediction base model is constructed through machine learning, including: Download a generalized model for flight price prediction based on supply and demand characteristics constructed through machine learning from the preset model library; Sorting the marginal supply and demand feature attributes for flight price relevance to obtain selected marginal supply and demand feature attributes; Based on the selected marginal supply and demand feature attributes and the core supply and demand feature attributes, transfer learning is performed on the supply and demand feature flight price prediction generalization model to obtain the supply and demand feature flight price prediction base model.

5. The method according to claim 4, wherein The steps for constructing the generalized model for flight price prediction based on supply and demand characteristics include: Based on the number of core supply and demand feature attributes, an equal number of fully connected branch neural networks are constructed; Build a fully connected neural network; Using the output of the fully connected branch neural network as the input of the fully connected neural network to construct a generalized model architecture for flight price prediction based on supply and demand characteristics; Taking the core supply and demand feature attributes as the only variable, the core supply and demand feature value dataset and the flight price identification dataset are collected, the supply and demand feature flight price prediction generalization model architecture is trained, and the supply and demand feature flight price prediction generalization model is obtained.

6. The method according to claim 5, wherein Based on the selected marginal supply and demand feature attributes and the core supply and demand feature attributes, transfer learning is performed on the supply and demand feature flight price prediction generalization model to obtain the supply and demand feature flight price prediction base model, including: Based on the number of selected edge supply and demand feature attributes, an equal number of incremental fully connected branch neural networks are constructed; Using the output of the incremental fully connected branch neural network as the input of the fully connected neural network to construct a supply and demand characteristic flight price prediction base model architecture; Freeze the fully connected branch neural network, use the selected edge supply and demand feature attribute as the only variable, collect the selected edge supply and demand feature value dataset and the flight price identification dataset, train the supply and demand feature flight price prediction base model architecture, and obtain a one-stage supply and demand feature flight price prediction base model; Activate the fully connected branch neural network, use the core supply and demand feature attributes and the selected edge supply and demand feature attributes as unique variables, collect supply and demand feature value datasets and flight price identification datasets, train the one-stage supply and demand feature flight price prediction base model, and obtain the supply and demand feature flight price prediction base model.

7. The method according to claim 1, wherein Retrieving the mode value of the flight price fluctuation vector of the historical flight sample set that meets the temporary feature structured array and setting it as the fluctuation curve correction vector, including: When the similarity between the temporary feature structured array of the historical sample and the temporary feature structured array is greater than or equal to the similarity threshold, the historical sample is added to the historical flight sample set; Performing a box plot analysis on the flight price fluctuation vector set of the historical flight sample set to obtain a box flight price fluctuation vector; The mean of the box flight price fluctuation vector is calculated and set as the fluctuation curve correction vector.

8. A machine learning dynamic flight information push system, characterized by: For implementing the method according to any one of claims 1 to 7, comprising: The data collection module is used to collect the supply and demand feature structured array, the time series feature structured array, and the temporary feature structured array based on the flight origin, flight destination, and demand time window input by the user, combined with the predefined supply and demand feature attributes, time series feature attributes, and temporary feature attributes; a price prediction module, configured to input the supply and demand feature structured array and the time series feature structured array into a flight price curve prediction model trained by machine learning to obtain an initial flight price fluctuation curve; a fluctuation correction module, configured to retrieve a mode value of a flight price fluctuation vector of a historical flight sample set that satisfies the temporary feature structured array, and set the mode value as a fluctuation curve correction vector; The sorting and pushing module is used to correct the initial flight price fluctuation curve according to the fluctuation curve correction vector, obtain an updated flight price fluctuation curve, sort the flight information from high to low according to the predicted price, and push it to the user end.

Images