運用深度學習與機器學習於旅店取消訂單量預測模型

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/93111

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/93111

Title:	運用深度學習與機器學習於旅店取消訂單量預測模型
Authors:	廖心妤;Liao, Hsin-Yu
Contributors:	資訊管理學系
Keywords:	旅店取消訂單;機器學習;深度學習;模型混合;hotel cancellation prediction;machine learning;deep learning;model hybriding
Date:	2023-06-30
Issue Date:	2024-09-19 16:42:33 (UTC+8)
Publisher:	國立中央大學
Abstract:	旅遊業是全球經濟中重要且快速發展的行業。近年來，旅遊業面臨著許多挑戰，直接影響到旅店業。特別是互聯網時代，線上訂房平台改變消費者行為，增加了取消訂單的頻率。為應對這些問題，旅店業需要準確的需求預測工具。本研究旨在建立旅店每日取消訂單量預測模型，並探索三種不同的方法在此領域的應用，包括分類模型、時間序列模型和模型混合。本研究使用個案公司提供的四間旅店真實資料集作為研究。在分類預測會使用機器學習算法，如Random Forest、XGBoost以及LGBM，將其應用在客戶訂單上；在時間序列預測則是使用簡單移動平均(SMA)、週期性移動平均(Periodic SMA)及LSTM Seq2Seq；模型混合方面是使用加權平均將分類和時間序列的預測結果做混合。結果顯示在四間旅店的分類模型實驗中，除客戶訂單特徵外，加入日期、時間相關特徵及特徵工程所創的新特徵能提高預測準確度。其次，時間序列模型實驗中使用單變量預測，深度學習模型(LSTM Seq2Seq)並不比傳統時間序列預測模型(Periodic SMA)準確。使用Periodic SMA於四間旅店的WMAPE平均下降約46%。此外，使用LSTM Seq2Seq，其多變量預測結果優於單變量，四間旅店的WMAPE平均下降約36%。其次，模型混合的結果比分類轉時序或時間序列模型更佳，四間旅店的WMAPE平均下降約3.3%。最後，四間旅店的數據聯合訓練並考慮旅店特徵，可進一步提高預測準確性，四間旅店的AUC平均提升約1.6%。由本研究的分析可知模型混合有助於為旅店業提供全面且有效的旅店取消訂單量預測解決方案。;Tourism is an important and fast-growing industry in the global economy. In recent years, the tourism industry has faced many challenges that directly affect the hotel industry. Especially in the Internet age, online room booking platforms have changed consumer behavior and increased the frequency of cancellations. To deal with these problems, the hotel industry needs accurate demand forecasting tools. The purpose of this study is to establish a forecasting model for daily cancellation orders in hotels, and to explore the application of three different methods in this field, including classification models, time series models, and model hybriding.This study uses the real data sets of four hotels provided by the case company as research. Machine learning algorithms, such as Random Forest, XGBoost, and LGBM, are used in classification forecasting and applied to customer orders; in time series forecasting, simple moving average (SMA), periodic moving average (Periodic SMA) and LSTM Seq2Seq are used ; In terms of model hybriding, the weighted average is used to hybrid the prediction results of classification and time series.The results show that: In the classification model experiments of the four hotels, in addition to customer order features, adding date and time-related features and new features created by feature engineering can improve the prediction accuracy; In the time series model experiments, single For variable prediction, the deep learning model (LSTM Seq2Seq) is not more accurate than the traditional time series (Periodic SMA). The WMAPE of the four hotels using Periodic SMA dropped by about 46% on average. In addition, using LSTM Seq2Seq, its multivariate prediction results are better than univariate, and the average WMAPE of the four hotels drops by about 36%. A decrease of about 3.3%;The joint training of the data of the four hotels and considering the characteristics of the hotels can further improve the prediction accuracy, and the AUC of the four hotels has increased by about 1.6% on average. The analysis of this study shows that model hybriding helps to provide a comprehensive and effective hotel cancellation order forecasting solution for the hotel industry.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	8	View/Open

社群 sharing

Loading...