DEEP LEARNING FOR HOURLY BUS BOARDING DEMAND PREDICTION FROM IMBALANCED SMART-CARD DATA

Abstract

Data from tap-on smart cards is a
useful tool for predicting future travel
demand and learning about customers'
boarding habits. In contrast to negative
instances (not boarding at that bus stop at that
time), positive instances (i.e., boarding at a
particular bus stop at a certain time) are
uncommon when looking at the smart-card
records (or instances) by boarding stops and
by time of day. It has been shown that
machine learning algorithms used to forecast
hourly boarding numbers from a certain site
are far less accurate when the data is
unbalanced. Prior to using the smart card data
to forecast bus boarding demand, this
research resolves the problem of data
imbalance. In order to supplement a synthetic
training dataset with more evenly distributed
travelling and non-traveling cases, we
suggest using deep generative adversarial
networks (Deep-GAN) to create fake
travelling instances. A deep neural network
(DNN) is then trained using the synthetic
dataset to forecast the travelling and nontraveling instances from a certain stop within
a specified time range. The findings
demonstrate that resolving the problem of
data imbalance may greatly enhance the
prediction model's functionality and better
match the true profile of passengers. A
comparison of the Deep-GAN's performance
with that of other conventional resampling
techniques demonstrates that the suggested
approach may generate a synthetic training
dataset with more variety and similarity, and
therefore, a better prediction capability. The
study emphasises the need of enhancing data
quality and model performance for travel
behaviour prediction and individual travel
behaviour analysis, and it offers helpful
advice in this regard