THESIS
2014
Abstract
With the development of modern technology, advanced GPS tracking solutions offer
us the best opportunity to learn the citys traffic knowledge. Travel cost on road segments
is the important hidden knowledge in the metropolitan city, which can be leveraged by
travel route planning, traffic event discovery, and citys fraud detection, etc. This paper
addresses the problem of travel cost estimation and prediction on road network with a not
only innovative but also practical solution. Although some previous work deal with the
same topic, they estimate travel cost on road segments based on the spatial and temporal
smoothness of adjacent neighbors, which is unreasonable under some circumstances. Other
works predicting travel cost ignore the obstacle of data sparsity by not predicting th...[
Read more ]
With the development of modern technology, advanced GPS tracking solutions offer
us the best opportunity to learn the citys traffic knowledge. Travel cost on road segments
is the important hidden knowledge in the metropolitan city, which can be leveraged by
travel route planning, traffic event discovery, and citys fraud detection, etc. This paper
addresses the problem of travel cost estimation and prediction on road network with a not
only innovative but also practical solution. Although some previous work deal with the
same topic, they estimate travel cost on road segments based on the spatial and temporal
smoothness of adjacent neighbors, which is unreasonable under some circumstances. Other
works predicting travel cost ignore the obstacle of data sparsity by not predicting the
travel cost of road segments without traffic data. In viewing this defect, we proposed the
methodology to estimate and predict travel cost based on objects similarity. To be more
specifically, we define the spatial temporal dividing sample as the minimal unit whose
travel cost is to be estimated or predicted. Knowledgeable samples present those who
have traffic data points, and their travel cost can be estimated by the traffic data. While
unknowledgeable samples are those who encounter the problem of data sparsity, and have
no estimation of travel cost since the lack of traffic data points. By extracting both the static and dynamic features for those samples, we profile them and apply the clustering
algorithm on them to identify similar samples. Within each cluster, we leverage the artificial
neural network to build the mapping relationship between knowledgeable samples
features and their travel cost. With the help of this mapping relationship, we finally infer
the clusters unknowledgeable samples travel cost. In terms of travel cost prediction, based
on the intuition that similar road segments share similar travel cost pattern, we put all
the road segments into different clusters and share the observed traffic data in the same
cluster to overcome the obstacle of data sparsity. And finally we leverage the time series
predicting model to predict the travel cost in the future. We evaluate our methodology
on one-month real traffic data from Shanghai. The experimental results on both small
data set and large data set show the validity and practicality of our methodology.
Post a Comment