Tourism demand prediction method based on two-stage feature selection and Time2Vec

Abstract

The invention relates to a tourism demand prediction method based on two-stage feature selection and Time2Vec, and belongs to the field of tourism data prediction. According to the method, a two-stage feature selection method based on MCC and ReliefF algorithms is designed, so that a most effective feature subset is selected from a data set; according to the method, an MCNN-BiLSTM model for predicting travel demands is provided, and intermediate layer information is fully utilized, so that important features are more effectively extracted from complex passenger flow volume related time sequence data; according to the method, Time2Vec is introduced, and a processing process of a time sequence is integrated into MCNN-BiLSTM, namely T2V-MCNN-BiLSTM, so that a signal processing algorithm commonly used in tourism demand prediction is improved, the model can efficiently extract advanced time features and predict tourism demands, and the prediction precision is improved.

Description

Technical Field

[0001] The invention belongs to the field of tourism data prediction, and specifically relates to a tourism demand prediction method based on two-stage feature selection and Time2Vec. Background Art

[0002] Tourism demand forecasting is of great significance to the development of tourism management departments and related industries. [1] Accurate demand forecasting can provide scientific basis and policy guidance for the strategic development, investment decision-making, revenue management and pricing strategies of tourism and related industries, thereby promoting the contribution of tourism to economic growth. [2,3] However, making accurate tourism demand forecasts is a difficult task. [4,5] , as customer traffic often varies seasonally and unpredictably [6] In addition, the impact of emergencies on people’s travel will also have a serious impact on the tourism industry. [7,8] Therefore, the complexity of passenger flow data and the ever-changing external environment make it a major challenge to achieve reliable forecasts.

[0003] Existing research can be divided into three categories:

[0004] (1) Time series models, which predict tourism demand based on historical patterns of tourism demand [9,10] ;

[0005] (2) Econometric models, which focus on establishing the causal relationship between tourism demand and various explanatory variables [11,12] ;

[0006] (3) Machine learning and deep learning models enable more accurate predictions by capturing latent patterns in high-dimensional feature spaces [13,14] .

[0007] The shortcomings of the existing technology are mainly reflected in:

[0008] (1) Existing tourism demand forecasting models fail to effectively deal with the uncertainty and complexity of external variables. Specifically, the ReliefF method, which is commonly used for feature selection, only eliminates redundant features by deleting features with smaller weights, which may lead to insufficient filtering of irrelevant features.

[0009] (2) Although existing signal processing methods for processing time series data improve prediction accuracy by reducing complexity, the decomposition results are unstable, which increases the complexity and uncertainty of data processing. In addition, separating the time series data decomposition and prediction processes may result in the loss of potentially important information during the decomposition process, affecting the subsequent feature extraction ability of the deep learning model.

[0010] (3) The existing tourism demand forecasting models do not fully utilize sample information during the training phase. That is, the CNN-LSTM hybrid model is combined in series, ignoring the large amount of useful information contained in the intermediate layers of the neural network.

[0011] References:

[0012] [1]Law, R., Li, G., Fong, DKC, & Han,

[0013] [2] Kaya, K., Y.,Yaslan,Y., G.,& F.(2022).Demandforecasting model using hotel clustering findings for hospitalityindustry.Information Processing&Management,59(1),102816.https: / / doi.org / 10.1016 / j.ipm.2021.102816.

[0014] [3]Wu,B.,Wang,L.,&Zeng,Y.-R.(2023).Interpretable tourism demandforecasting with temporal fusion transformers amid COVID-19.AppliedIntelligence,53(11),14493–14514.https: / / doi.org / 10.1007 / s10489-022-04254-0.

[0015] [4]Silva,E.S.,Hassani,H.,Heravi,S.,&Huang,X.(2019).Forecastingtourism demand with denoised neural networks.Annals of Tourism Research,74,134–154.https: / / doi.org / 10.1016 / j.annals.2018.11.006.

[0016] [5]Song,H.,Qiu,R.T.,&Park,J.(2019).A review ofresearch on tourismdemand forecasting:Launching the Annals of Tourism Research CuratedCollection on tourism demand forecasting.Annals ofTourism Research,75,338–362.https: / / doi.org / 10.1016 / j.annals.2018.12.001.

[0017] [6]Zhang,C.,&Tian,Y.X.(2022).Forecast daily tourist volumes duringthe epidemic period using COVID-19 data,search engine data and weatherdata.Expert Systems with Applications,210,Article 118505.https: / / doi.org / 10.1016 / j.eswa.2022.118505.

[0018] [7] M.,Soriano,D.R.,&Porada-Rochoń,M.(2021).Impact of COVID-19 onthe travel and tourism industry.Technological forecasting and social change,163,120469.https: / / doi.org / 10.1016 / j.techfore.2020.120469.

[0019] [8]Xu,S.,Liu,Y.,&Jin,C.(2023).Forecasting daily tourism demand withmultiple factors.Annals ofTourism Research.https: / / doi.org / 10.1016 / j.annals.2023.103675.

[0020] [9]Song,H.,Qiu,R.T.,&Park,J.(2019).A review ofresearch on tourismdemand forecasting:Launching the Annals of Tourism Research CuratedCollection on tourism demand forecasting.Annals ofTourism Research,75,338–362.https: / / doi.org / 10.101.

[0021]

[10] Park,S.,Lee,J.,&Song,W.(2017).Short-term forecasting ofJapanesetourist inflow to South Korea using google trends data.Journal of Travel&Tourism Marketing,34(3),357–368.https: / / doi.org / 10.1080 / 10548408.2016.11706516 / j.annals.2018.12.001.

[0022]

[11] Cao,Z.,Li,G.,&Song,H.(2017).Modelling the interdependence oftourism demand:The global vector autoregressive approach.Annals of TourismResearch,1–13.https: / / doi.org / 10.1016 / j.annals.2017.07.019.

[0023]

[12] Li, G., Wong, KKF, Song, H., & Witt, SF (2006). Tourism Demand Forecasting: ATime Varying Parameter Error Correction Model. Journal of Travel Research, 45(2), 175–185. https: / / doi.org / 10.1177 / 0047287506291596.

[0024]

[13] Polyzos, S., Samitas, A., & Spyridou, A. Ef. (2021). Tourism demand and the COVID-19 pandemic: an LSTM approach. Tourism Recreation Research, 175–187. https: / / doi.org / 10.1080 / 02508281.2020.1777053.

[0025]

[14] Zhang, C., & Tian, ​​YX (2022). Forecast daily tourist volumes during the period epidemic using COVID-19data, search engine data and weatherdata. Expert Systems with Applications, 210, Article 118505. https: / / doi.org / 10.1016 / j.eswa.2022.118505. Summary of the invention

[0026] The purpose of the present invention is to overcome the shortcomings of the prior art, provide a tourism demand forecasting method based on two-stage feature selection and Time2Vec, design a two-stage feature selection method based on MCC and ReliefF algorithms to select the most effective feature subset from a data set; propose a MCNN-BiLSTM model for predicting tourism demand, make full use of the intermediate layer information, so as to more effectively extract important features from complex passenger flow related time series data; introduce Time2Vec, integrate the time series processing process into MCNN-BiLSTM, namely T2V-MCNN-BiLSTM, thereby improving the signal processing algorithm commonly used in tourism demand forecasting, so that the model can efficiently extract high-level time features, predict tourism demand, and thus improve the prediction accuracy.

[0027] To achieve the above purpose, the technical solution of the present invention is: a tourism demand forecasting method based on two-stage feature selection and Time2Vec, comprising:

[0028] Design a two-stage feature selection method based on MCC and ReliefF algorithms to select a subset of key features from the dataset;

[0029] A MCNN-BiLSTM model for predicting tourism demand is proposed, and Time2Vec is introduced to integrate the processing of time series into the MCNN-BiLSTM model to form a T2V-MCNN-BiLSTM model to predict tourism demand based on a subset of key features.

[0030] In one embodiment of the present invention, the data set is obtained as follows:

[0031] (1) Collect historical passenger flow data: The historical passenger flow of the scenic spot is collected from the official website of the scenic spot;

[0032] (2) Collect Baidu Index data: Use Python programming to obtain Baidu Index data for scenic spot-related search keywords. The strategy for selecting keywords is to use the core words related to tourist attractions and use the “related search” strategy provided by the Baidu Index data platform to induce other related keywords or phrases;

[0033] (3) Collect emergency data: collect daily updates on emergency-related indicators;

[0034] (4) Collect other data: including weather data and holiday data. Weather data includes weather conditions and temperature, which are collected from the weather network.

[0035] In one embodiment of the present invention, before using the data set, it is necessary to remove duplicate and invalid data in the data set, use interpolation to complete the data, and then perform normalization processing.

[0036] In one embodiment of the present invention, a two-stage feature selection method based on MCC and ReliefF algorithms is used to select a key feature subset from a data set by calculating the MIC of each variable of order 0-n to determine the optimal lag structure, setting the maximum lag order to 30, arranging the time series according to the lag order, and determining the key features using a two-stage feature selection method to obtain a key feature subset.

[0037] In one embodiment of the present invention, the implementation method of predicting tourism demand based on the key feature subset through the T2V-MCNN-BiLSTM model is as follows:

[0038] (1) Extracting time embedding features: Using Time2Vec to automatically capture the periodic and non-periodic time features in the tourism demand-related time series in the key feature subset; after obtaining the time embedding, it is connected with the original feature data in the key feature subset as the input of the MCNN-BiLSTM model;

[0039] (2) Extracting time and variable interaction features: In the MCNN-BiLSTM model, CNN is used to extract features with high information content and high correlation in all input vectors, BiLSTM is used to learn the features extracted by CNN to obtain more contextual information, and a connection function is introduced to fuse the features extracted by CNN and BiLSTM;

[0040] (3) Output prediction results: The fused features are processed through the fully connected layer to obtain the final prediction results.

[0041] In one embodiment of the present invention, the two-stage feature selection method is specifically as follows:

[0042] In the first stage, MCC(θ) is used to calculate the amount of repeated information between each feature in the key feature subset and other features to filter out some redundant features; the features are filtered according to the custom threshold β, and the features are retained when θ is greater than β; the mth eigenvalue of the nth sample is calculated using f nm Indicates that the mean of the i-th feature is defined as Covariance v ij and the correlation coefficient r ij Defined as:

[0043]

[0044] The covariance matrix and correlation coefficient matrix are defined as V = {v ij} and R = {r ij}, where i,j∈(1,2,…,m);

[0045] MCC is calculated based on the correlation coefficient matrix R:

[0046]

[0047] where θ i is the MCC between the ith feature and other features, r i 、r i ' and R i-1 They are respectively represented as the correlation coefficient matrix of the i-th feature, r i The transpose of and the correlation coefficient matrix of other features except the i-th feature;

[0048] The second stage is to determine the final features using the ReliefF algorithm based on the features screened in the first stage;

[0049] First, randomly select a sample S, and then select the q nearest neighbors H of the same type as the sample S. q ={h1,h2,h3,...,h q}; In addition, find the q nearest neighbors of different types G q ={g1,g2,g3,...,g q}, Gq type is C, and the weight of each feature F is calculated as follows:

[0050]

[0051] Where diff(F,S,s) represents the difference between samples S and s on feature F, S(F) and s(F) are the numerical values ​​of feature F of samples S and s respectively, P(C) and P(class(P)) are the prior probabilities of type C and randomly selected sample types respectively, that is, the proportion of samples of this type in the total number of samples, max(F) and min(F) are the maximum and minimum values ​​of feature F, and m is the number of iterations.

[0052] In one embodiment of the present invention, the T2V-MCNN-BiLSTM model is implemented as follows:

[0053] (1)Time2Vec

[0054] For a scalar concept of a given time τ, the Time2Vec formula for the i-th element is as follows:

[0055]

[0056] Among them, ω i and is a learnable parameter, and F is a periodic activation function;

[0057] In general, given a time series, Time2Vec encodes it into a vector of size k;

[0058] (2) MCNN-BiLSTM

[0059] The CNN-BiLSTM model combines CNN and BiLSTM. First, CNN can extract more abstract and information-rich feature representations from the original time series data. Second, BiLSTM, as the lower-level component in the CNN-BiLSTM model, receives the time information of important features of tourism demand prediction proposed by CNN. BiLSTM combines information flow in two directions, namely forward and backward, allowing the model to consider past and future information while predicting each time step, thereby more comprehensively understanding the temporal relationship in the sequence, reducing information loss and overcoming the gradient vanishing problem in long sequences. The outputs of the forward and backward layers of BiLSTM are connected together to form the final output. The relevant calculation process is as follows:

[0060]

[0061]

[0062] Where LSTM stands for long short-term memory network, x i is the input vector of the input sequence at time step i, are the hidden states of the forward LSTM and the backward LSTM at time step i, respectively. [:,:] represents the concatenation operation of the vector;

[0063] An additional flattening layer is added after each pooling layer of CNN to make full use of the intermediate layer information, and the pooling layer selects the maximum pooling layer.

[0064] In one embodiment of the present invention, the periodic activation function uses a sine function sin.

[0065] The present invention also provides a tourism demand forecasting system based on two-stage feature selection and Time2Vec, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0066] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] (1) By fusing the features extracted by the CNN intermediate layer and BiLSTM, the present invention can better extract usable information from multi-source data, thereby achieving more accurate prediction.

[0069] (2) The present invention utilizes a two-stage feature selection strategy to screen a limited number of features, which makes the model more stable and effectively solves the dimensionality disaster caused by too many external variables. The experimental results prove the effectiveness of our method.

[0070] (3) The present invention uses Time2Vec to replace the cumbersome time series data decomposition process and merges it into a hybrid deep learning model composed of different deep neural networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is the model framework of the method of the present invention.

[0072] Figure 2 This is a tourist demand forecast result of a scenic spot according to an example of the present invention. DETAILED DESCRIPTION

[0073] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0074] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0075] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0076] The present invention provides a tourism demand forecasting method based on two-stage feature selection and Time2Vec, comprising:

[0077] Design a two-stage feature selection method based on MCC and ReliefF algorithms to select a subset of key features from the dataset;

[0078] A MCNN-BiLSTM model for predicting tourism demand is proposed, and Time2Vec is introduced to integrate the processing of time series into the MCNN-BiLSTM model to form a T2V-MCNN-BiLSTM model to predict tourism demand based on a subset of key features.

[0079] The following is the specific implementation process of the present invention.

[0080] This embodiment provides a tourism demand forecasting method based on two-stage feature selection and Time2Vec. The basic framework of the model is shown in the attached figure. Figure 1 , including the following steps:

[0081] The first step is data collection and preprocessing:

[0082] The prediction model uses four types of data, namely historical passenger flow data, Baidu Index data, emergency data, and other data. The data collection methods are as follows: (1) Collect historical passenger flow data: The historical passenger flow of the scenic spot is collected from the official website of the scenic spot; (2) Collect Baidu Index data: Use Python programming to obtain Baidu Index data of relevant search keywords of the scenic spot. The strategy for selecting keywords is to use the central words related to tourist attractions and use the "related search" strategy provided by the Baidu Index platform to induce other related keywords or phrases; (3) Collect emergency data: Collect daily new indicators related to emergency events through relevant official channels; (4) Collect other data: Other data include weather data, holiday data, etc., among which weather data includes weather conditions and temperature, which are collected from the weather website (https: / / lishi.tianqi.com / ).

[0083] The data preprocessing method is as follows: remove duplicate and invalid data from historical passenger flow, emergency events and other data, and use interpolation to complete the data.

[0084] The second step is feature processing

[0085] (1) Data normalization: Data such as weather and holidays are processed into data that the model can learn and understand. The maximum and minimum normalization method is used to normalize continuous data to the range of [0, 1], and one-hot encoding is used to process discrete data.

[0086] (2) Select features using a two-stage feature selection method: Calculate the MIC of each variable from 0 to n orders to determine the optimal lag structure, set the maximum lag order to 30, arrange the time series according to the lag order, and use a two-stage feature selection method to determine the key features.

[0087] The third step is to predict tourism demand based on the T2V-MCNN-BiLSTM model.

[0088] Input the processed data into the T2V-MCNN-BiLSTM hybrid model to obtain the corresponding output value. The specific steps are as follows:

[0089] (1) Extracting time embedding features: The Time2Vec layer is used to automatically capture important periodic and non-periodic time features in the tourism demand-related time series. After obtaining the time embedding, it is concatenated with the original feature data as the input of the MCNN-BiLSTM.

[0090] (2) Extracting temporal and variable interaction features: In the MCNN-BiLSTM layer, the superiority of CNN feature extraction is utilized to extract features with high information content and high correlation in all input vectors. The features extracted by CNN are learned using BiLSTM to obtain more contextual information, and a connection function is introduced to effectively fuse the features extracted by the CNN intermediate layer and BiLSTM.

[0091] (3) Output prediction results: The fused data is processed through the fully connected layer to obtain the final prediction results. The model architecture of T2V-MCNN-BiLSTM is shown in the attached Figure 1 (right).

[0092] In the above steps, the two-stage feature selection method and the T2V-MCNN-BiLSTM model are the focus of the present invention, which are discussed in detail below.

[0093] 1. Two-stage feature selection method

[0094] The two-stage feature selection method consists of two parts: multivariate correlation coefficient (MCC) and ReliefF algorithm.

[0095] In the first stage, MCC(θ) is used to calculate the amount of repeated information between each feature in the key feature subset and other features to filter out some redundant features; the features are filtered according to the custom threshold β, and the features are retained when θ is greater than β; the mth eigenvalue of the nth sample is calculated using f nm Indicates that the mean of the i-th feature is defined as Covariance v ij and the correlation coefficient r ij Defined as:

[0096]

[0097] The covariance matrix and correlation coefficient matrix are defined as V = {v ij} and R = {r ij}, where i,j∈(1,2,…,m). MCC is calculated based on the correlation coefficient matrix R:

[0098]

[0099] where θ i is the MCC between the ith feature and other features, r i 、r i' and R i-1 They are respectively represented as the correlation coefficient matrix of the i-th feature, r i The transpose of and the correlation coefficient matrix of other features except the i-th feature;

[0100] The second stage is to determine the final features using the ReliefF algorithm based on the features screened in the first stage;

[0101] First, randomly select a sample S, and then select the q nearest neighbors H of the same type as the sample S. q ={h1,h2,h3,...,h q}; In addition, find the q nearest neighbors of different types G q ={g1,g2,g3,...,g q}, Gq type is C, and the weight of each feature F is calculated as follows:

[0102]

[0103]

[0104] Where diff(F,S,s) represents the difference between samples S and s on feature F, S(F) and s(F) are the numerical values ​​of sample S and s on feature F, respectively. P(C) and P(class(P)) are the prior probabilities of type C and randomly selected sample types, that is, the proportion of samples of this type in the total number of samples, max(F) and min(F) are the maximum and minimum values ​​of feature F, and m is the number of iterations.

[0105] 2. T2V-MCNN-BiLSTM model

[0106] The T2V-MCNN-BiLSTM model consists of two components:

[0107] (1)Time2Vec

[0108] Time2Vec is a time vector representation method that can effectively map discrete time information into a continuous space, which helps the model learn temporal dependencies, such as seasonal changes, trends, and periodicity (Kazemi, S. et.al, 2019). For a scalar concept of a given time τ, the Time2Vec formula for the i-th element is as follows:

[0109]

[0110] Among them, ω i and is a learnable parameter, F is a periodic activation function, and we use the sine function as the activation function.

[0111] In general, given a time series, Time2Vec encodes it into a vector of size k. This adaptive time embedding representation can be learned by other modules of the deep learning model, so we apply Time2Vec to our model. By adding Time2Vec, the proposed deep learning model can dynamically extract temporal features from the time series without the need for cumbersome time series processing steps such as EMD, which can further improve the accuracy and efficiency of the prediction model.

[0112] (2) MCNN-BiLSTM

[0113] The CNN-BiLSTM model combines CNN and BiLSTM to fully utilize the advantages of CNN in feature extraction and the powerful capabilities of BiLSTM in time series modeling. First, CNN can extract more abstract and information-rich feature representations from the original time series data. Because of its local connectivity and global sharing characteristics, the network can capture temporal dependencies and correlations between features related to tourism demand prediction, improving the accuracy and generalization of the model. Secondly, BiLSTM, as the lower-level component in the CNN-BiLSTM architecture, receives the temporal information of important features of tourism demand prediction proposed by CNN. By combining information flow in two directions (forward and backward), BiLSTM allows the model to consider past and future information while predicting each time step, thereby more comprehensively understanding the temporal relationship in the sequence, reducing information loss and overcoming the gradient vanishing problem in long sequences. The outputs of the forward and backward layers of BiLSTM are connected together to form the final output. The relevant calculation process is as follows:

[0114]

[0115] LSTM stands for Long Short-Term Memory, which is a variant of Recurrent Neural Network (RNN) that can effectively process and memorize long-distance dependencies in sequence data. i is the input vector of the input sequence at time step i, are the hidden states of the forward LSTM and backward LSTM at time step i respectively, and [:,:] represents the concatenation operation of the vector.

[0116] Different from the CNN-BiLSTM hybrid model, T2V-MCNN-BiLSTM adds an extra flattening layer after each pooling layer of CNN to make full use of the information in the intermediate layers. In time series prediction tasks, the performance of the maximum pooling operation is usually better than the average pooling operation, so the maximum pooling layer is selected.

[0117] First, the Time2Vec layer is used to automatically capture the important periodic and non-periodic time features in the tourism demand-related time series. After obtaining the time embedding, it is connected with the original feature data as the input of MCNN-BiLSTM. Then, in the MCNN-BiLSTM layer, the advantages of CNN feature extraction are used to extract the features with high information content and high correlation in all input vectors, and the features extracted by CNN are learned using BiLSTM to obtain more contextual information, and the connection function is introduced to effectively fuse the features extracted by the CNN intermediate layer and BiLSTM. Finally, the fused data is processed by the fully connected layer to obtain the final prediction result.

[0118] Example

[0119] Based on the daily tourist volume data of two famous domestic scenic spots, the present invention analyzes the effectiveness of the hybrid deep learning model based on Time2Vec in tourism demand prediction.

[0120] The present invention first uses a two-stage feature selection method to calculate the top 3 features as the optimal subset. Before model training, it is necessary to first determine the important hyperparameters of the model, including the kernel size of CNN, the number of units of BiLSTM, etc. Taking the three-step prediction of one of the scenic spots as an example, the optimal parameter combination is 2 CNN layers, the number of filters and convolution kernel size of the first layer CNN are 32 and 2, the number of filters and convolution kernel size of the second layer CNN are 64 and 2, the activation function is ReLU, the number of units of BiLSTM is 64, and the number of neurons in the fully connected layer is 32.

[0121] In order to prove the effectiveness and superiority of the proposed model in tourism demand forecasting, we conducted a large number of experiments. Experiment 1 is an ablation experiment, which aims to prove the necessity of each component of the model. Experiment 2 compares the performance with some baseline models to prove the superiority of our proposed model. The baseline models are as follows:

[0122] SVR: Support vector regression,

[0123] ETS: exponential smoothing

[0124] ARIMAX: autoregressive integrated moving average with explanatory variable,

[0125] SARIMAX: seasonal autoregressive integrated moving average with explanatory variable,

[0126] GRU: Gated recurrent unit networks,

[0127] LSTM-AM: LSTM with the attention mechanism.

[0128] Some of the results are shown in Appendix 1, Appendix 2, and Appendix Figure 2 shown.

[0129] Appendix 1 Comparison of prediction performance of each model in Experiment 1

[0130]

[0131] Appendix 2 Prediction performance of different scenes in scenic spot 1

[0132]

[0133] Experiment 1: In order to verify the effectiveness of each module, the present invention designed several models for ablation experiments, and analyzed the influence and contribution of each module of the model network by removing specific modules. First, in order to verify the effectiveness of each component of MCNN-BiLSTM, four models, including CNN, BiLSTM, CNN-BiLSTM, and MCNN-BiLSTM, were designed. Then, in order to verify the impact of Time2Vec on model performance, three models, including T2V-CNN, T2V-BiLSTM, and T2V-MCNN-BiLSTM, were designed. Finally, we compared the performance of three models, including original data (T2V-MCNN-BiLSTM-O), PCA feature selection (T2V-MCNN-BiLSTM-P), and ReliefF feature selection (T2V-MCNN-BiLSTM-R), with the two-stage feature selection model we proposed. Appendix 1 shows the MAE, RMSE, and MAPE results obtained from the 10 model experiments, and the best model evaluation indicators are indicated in bold. It should be noted that for other models, we use the data set after two-stage feature selection. The results show that the MCNN-BiLSTM, Time2Vec layer, and two-stage feature selection method in the hybrid model all help to improve the model performance, and our proposed hybrid model performs the best.

[0134] Experiment 2: In order to comprehensively evaluate the predictive effectiveness and superiority of the proposed model, we conducted single-step and multi-step prediction experiments on the scenic spot 1 dataset. Appendix 2 shows the prediction performance of different scenarios in scenic spot 1. The results show that the model T2V-MCNN-BiLSTM proposed in this invention performs best. The traditional models ETS and SVR cannot capture the complex nonlinear patterns and dynamic changes in time series data, so the model prediction results are the worst. Compared with other models, the T2V-MCNN-BiLSTM model has an average improvement of 44.75%, 50.56% and 49.79% in MAPE of single-step, 3-step and 7-step predictions. Taking the 3-step prediction as an example, Appendix Figure 2 It is the visualization result of different models for predicting the tourism demand of scenic spot 1, which provides a more intuitive performance comparison. The predicted value of T2V-MCNN-BiLSTM is closest to the actual value, which proves the effectiveness of our model.

[0135] The present invention also provides a tourism demand forecasting system based on two-stage feature selection and Time2Vec, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0136] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0137] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0138] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, 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 processor, or other programmable data processing device to generate 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 flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.

[0139] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0140] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0141] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A tourism demand forecasting method based on two-stage feature selection and Time2Vec, characterized in that: include: Design a two-stage feature selection method based on MCC and ReliefF algorithms to select a subset of key features from the dataset; A MCNN-BiLSTM model for predicting tourism demand is proposed, and Time2Vec is introduced to integrate the processing of time series into the MCNN-BiLSTM model to form a T2V-MCNN-BiLSTM model to predict tourism demand based on a subset of key features.

2. The tourism demand forecasting method based on two-stage feature selection and Time2Vec according to claim 1 is characterized in that: The dataset is obtained as follows: (1) Collect historical passenger flow data: The historical passenger flow of the scenic spot is collected from the official website of the scenic spot; (2) Collect Baidu Index data: Use Python programming to obtain Baidu Index data for scenic spot-related search keywords. The strategy for selecting keywords is to use the core words related to tourist attractions and use the "related search" strategy provided by the Baidu Index data platform to induce other related keywords or phrases; (3) Collect emergency data: collect daily updates on emergency-related indicators; (4) Collect other data: including weather data and holiday data. Weather data includes weather conditions and temperature, which are collected from the weather network.

3. The tourism demand forecasting method based on two-stage feature selection and Time2Vec according to claim 1 is characterized in that: Before using the dataset, it is necessary to remove duplicate and invalid data in the dataset, use interpolation to complete the data, and then perform normalization.

4. The tourism demand forecasting method based on two-stage feature selection and Time2Vec according to claim 1 is characterized in that: The two-stage feature selection method based on MCC and ReliefF algorithms selects the key feature subset from the data set by calculating the MIC of each variable from order 0 to n to determine the optimal lag structure, setting the maximum lag order to 30, arranging the time series according to the lag order, and determining the key features using the two-stage feature selection method to obtain the key feature subset.

5. The tourism demand forecasting method based on two-stage feature selection and Time2Vec according to claim 1 is characterized in that: Through the T2V-MCNN-BiLSTM model, based on the key feature subset, the implementation method of predicting tourism demand is as follows: (1) Extracting temporal embedding features: Time2Vec is used to automatically capture the periodic and non-periodic time features in the tourism demand-related time series in the key feature subset; After obtaining the temporal embedding, it is concatenated with the original feature data in the key feature subset as the input of the MCNN-BiLSTM model; (2) Extracting time and variable interaction features: In the MCNN-BiLSTM model, CNN is used to extract features with high information content and high correlation in all input vectors, BiLSTM is used to learn the features extracted by CNN to obtain more contextual information, and a connection function is introduced to fuse the features extracted by CNN and BiLSTM; (3) Output prediction results: The fused features are processed through the fully connected layer to obtain the final prediction results.

6. A tourism demand forecasting method based on two-stage feature selection and Time2Vec according to claim 1 or 4, characterized in that: The two-stage feature selection method is as follows: In the first stage, MCC(θ) is used to calculate the amount of repeated information between each feature in the key feature subset and other features to filter out some redundant features; the features are filtered according to the custom threshold β, and the features are retained when θ is greater than β; the mth eigenvalue of the nth sample is calculated using f nm Indicates that the mean of the i-th feature is defined as Covariance v ij and the correlation coefficient r ij Defined as: The covariance matrix and correlation coefficient matrix are defined as V = {v ij } and R = {r ij }, where i,j∈(1,2,…,m); MCC is calculated based on the correlation coefficient matrix R: where θ i is the MCC between the ith feature and other features, r i 、r i ' and R i-1 They are respectively represented as the correlation coefficient matrix of the i-th feature, r i The transpose of and the correlation coefficient matrix of other features except the i-th feature; The second stage is to determine the final features using the ReliefF algorithm based on the features screened in the first stage; First, randomly select a sample S, and then select the q nearest neighbors H of the same type as the sample S. q ={h1,h2,h3,...,h q }; In addition, find the q nearest neighbors of different types G q ={g1,g2,g3,...,g q }, Gq type is C, and the weight of each feature F is calculated as follows: Where diff(F,S,s) represents the difference between samples S and s on feature F, S(F) and s(F) are the numerical values ​​of feature F of samples S and s respectively, P(C) and P(class(P)) are the prior probabilities of type C and randomly selected sample types respectively, that is, the proportion of samples of this type in the total number of samples, max(F) and min(F) are the maximum and minimum values ​​of feature F, and m is the number of iterations.

7. A tourism demand forecasting method based on two-stage feature selection and Time2Vec according to claim 1 or 5, characterized in that: The T2V-MCNN-BiLSTM model is implemented as follows: (1)Time2Vec For a scalar concept of a given time τ, the Time2Vec formula for the i-th element is as follows: Among them, ω i and is a learnable parameter, and F is a periodic activation function; In general, given a time series, Time2Vec encodes it into a vector of size k; (2) MCNN-BiLSTM The CNN-BiLSTM model combines CNN and BiLSTM. First, CNN can extract more abstract and information-rich feature representations from the original time series data. Second, BiLSTM, as the lower-level component in the CNN-BiLSTM model, receives the time information of important features of tourism demand prediction proposed by CNN. BiLSTM combines information flow in two directions, namely forward and backward, allowing the model to consider past and future information while predicting each time step, thereby more comprehensively understanding the temporal relationship in the sequence, reducing information loss and overcoming the gradient vanishing problem in long sequences. The outputs of the forward and backward layers of BiLSTM are connected together to form the final output. The relevant calculation process is as follows: Where LSTM stands for long short-term memory network, x i is the input vector of the input sequence at time step i, are the hidden states of the forward LSTM and the backward LSTM at time step i, respectively. [:,:] represents the concatenation operation of the vector; An additional flattening layer is added after each pooling layer of CNN to make full use of the intermediate layer information, and the pooling layer selects the maximum pooling layer.

8. The tourism demand forecasting method based on two-stage feature selection and Time2Vec according to claim 7 is characterized in that: The periodic activation function uses the sine function sin.

9. A tourism demand forecasting system based on two-stage feature selection and Time2Vec, characterized in that: The method comprises a memory, a processor and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps as claimed in any one of claims 1 to 8 can be implemented.

10. A computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, and when the processor executes the computer program instructions, the method steps according to any one of claims 1 to 8 can be implemented.

Images