基于TPA-Transformer的机票价格预测  

Airfare Price Prediction Based on TPA-Transformer

作　　者：申志豪 李娜 尹世豪 杜一[1,2,3] 胡良霖 SHEN Zhihao;LI Na;YIN Shihao;DU Yi;HU Lianglin(Computer Network Information Center,Chinese Academy of Sciences,Beijing 100083,China;University of China Academy of Sciences,Beijing 100049,China;National Basic Science Data Center,Beijing 100083,China;Travelsky Technology Limited,Beijing 101318,China;Key Laboratory of Intelligent Passenger Service of Civil Aviation,Beijing 101318,China)

机构地区：[1]中国科学院计算机网络信息中心,北京100083 [2]中国科学院大学,北京100049 [3]国家基础学科公共科学数据中心,北京100083 [4]中国民航信息网络股份有限公司,北京101318 [5]民航旅客服务智能化应用技术重点实验室,北京101318

出　　处：《数据与计算发展前沿》2023年第6期115-125,共11页Frontiers of Data & Computing

基　　金：国家自然科学基金重点项目(61836013);北京市科技新星计划(Z191100001119090)。

摘　　要：【目的】航空业被认为是使用最复杂定价策略的行业之一,机票价格动态波动显著。乘客根据经验购票(如尽早购票)往往得不到最优价格,而基于时序模型的传统价格预测方法也不能很好地捕捉复杂内外部因素与机票价格之间的依赖关系。【方法】本文提出基于机器学习的机票价格预测模型TPA-Transformer(Ticket Price Aware Transformer,机票价格感知Transformer)和一种基于时间序列的数据处理方法,通过增加注意力模块引入其他航班价格参考信息,并在Encoder后增加多层卷积结构进行多航班不同属性信息融合与局部特征提取,从而提高模型在多步价格预测上的表现性能。【结果】随后在5个回归评价指标(MSE、RMSE、MAE、ACC以及AMS)上对模型结果进行验证。【结论】结果证明,模型能够有效提高预测准确率,明显优于其他5个对比模型(随机森林、XGBoost、LSTM、GRU、Transformer)。[Objective]The airline industry is one of the industries that use the most complex pricing strat-egies.Even the ticket prices on the same flight fluctuate dynamically and significantly.Passen-gers can only make decisions based on experience,such as buying tickets as soon as possible,but it is not always the best choice.The traditional price forecasting methods cannot sufficient-ly capture the dependence between complex internal/external factors and the ticket prices.[Methods]This paper designs and implements the TPA-Transformer(Ticket Price Aware Transformer)to predict ticket price and proposes a related data processing method based on time series.This model adds the attention module to introduce reference information and the multi-layer convolution after the encoder to fuse the informa-tion of different features of multiple flights and extract local features to improve the model’s performance in multi-step price prediction.[Results]The model is verified on five regression evaluation indexes(MSE,RMSE,MAE,ACC,and AMS).[Conclusions]Experiments show that the model effectively improves the prediction accura-cy and is superior to other five comparison models(Random Forest,XGBoost,LSTM,GRU,and Transformer).

关 键 词：机票价格预测 机器学习 时间序列 注意力机制 

分 类 号：F562[经济管理—产业经济] TP18[自动化与计算机技术—控制理论与控制工程]