摘要
航班延误是引起乘客投诉的主要原因.针对航班延误预测问题,以一维卷积神经网络为基础,提出了一个航班延误预测网络.该网络融合了多种影响航班延误的相关性因素,利用金字塔池化层以适应不同长度的样本数据,同时引入密集连接结构及注意力模块对网络进行改进以提升准确率和精确率.针对数据不平衡导致的召回率偏低的问题,从算法和数据两个角度进行优化.在算法角度,使用代价敏感损失函数对不平衡数据和难易样本进行权重平衡;在数据角度,采用生成对抗网络对延误样本做数据增强以平衡航班数据.实验结果表明:相对于基准网络,本文模型的F1值提升了8.5%.使用代价敏感损失函数后,模型的准确率得到了小幅提升,而模型的召回率提升了1.6%.同时,使用深度生成网络平衡航班数据后,模型的召回率提升了15.8%.
Flight delay is the main cause of passenger complaints.In order to solve the problem,a flight delay prediction network is proposed based on a one-dimensional convolutional neural network.It integrates a variety of correlation factors affecting flight delays,adopts a pyramid pooling layer to adapt to data samples of different lengths,and introduces dense connection structures and attention modules to improve the network.This paper discusses the problem of low recall rates caused by data imbalance from the perspective of algorithms and data.In terms of algorithms,it uses a cost-sensitive loss function to balance the weight of unbalanced data and difficult samples;from the data point of view,it uses generative adversarial networks to enhance the delayed samples to balance flight data.Experimental results show that compared with the benchmark network,F1 value of the proposed model is improved by 8.5%.After using the cost-sensitive loss function,the model's accuracy slightly improved,while the model's recall is increased by 1.6%.Besides that,after using the deep generative network to balance the flight data,the model's recall is improved by 15.8% compared with the original model.
作者
王冬雪
闻翔
孙慧妮
蒋云鹏
白双
WANG Dongxue;WEN Xiang;SUN Huini;JIANG Yunpeng;BAI Shuang(School of Electronic and Information Engineering,Beijing Jiaotong University,Beijing 100044,China;Big Data Analysis and Application Center,China Academy of Civil Aviation Science and Technology,Beijing 100028,China)
出处
《北京交通大学学报》
CAS
CSCD
北大核心
2023年第2期95-105,共11页
JOURNAL OF BEIJING JIAOTONG UNIVERSITY
基金
国家自然科学基金(61602027)。
关键词
航班延误预测
航班相关效应
一维卷积神经网络
不均衡数据
代价敏感学习
深度生成网络
flight delay prediction
correlation effects on flights
one-dimensional convolutional neural network
imbalanced data
cost-sensitive learning
deep generation network
