September 24, 2019
Six Questions on Network-wide Traffic Prediction
College of Metropolitan Transportation, Beijing University of Technology, China
Recently, many network-wide traffic prediction studies based on various deep-learning
methods are published. The field becomes very active, and the relevant papers are getting
many attractions and “crazy” citations. After reading and reviewing many related papers, I
have many doubts and cannot help asking the following six questions.
1. Is the proposed model innovative?
Many studies make network-wide prediction by directly employing the existing machine
learning models or assembling several existing models (Figure 1). Very few deeply revises the
structure of the machine learning models, not to mention proposing a completely new model
to make prediction (After all, traffic prediction is like an application field of computer science,
sometimes). Therefore, do such studies have significant originality and deserve being
published in top-ranking transportation journals?
Figure 1. Is LEGO innovative or the figure made by LEGO pieces?
September 24, 2019
2. Is the heterogeneity of road segments considered?
Commonly, an existing map (such as OpenStreetMap) is employed for map-matching and
then prediction. Do people really care about the road segmentation? As shown in Figure 2,
some segments are long while some are really short. After map-matching GPS data or loop
detector data, traffic conditions can be estimated and obtained. However, it is obvious that
the weights of the segments are different in the final result, as well as the accuracy of using
instantaneous traffic data to represent traffic conditions on a road segment. Therefore, is it
problematic to directly use those existing maps and ignore how those roads are segmented?
Figure 2. Is it reasonable to treat long and short road segment to be the same?
3. Is the unevenness of traffic conditions considered?
It is known from the traffic flow theory that the easiest part of prediction is free-flow and
jammed traffic, while the most difficult part is the traffic at capacity and slightly congested
traffic. When summing up the prediction accuracy on roads to obtain a global metric, the
percentage of each traffic condition will be crucial to the final result. Therefore, attentions
must be paid when using a global metric.
4. Is the competitive method sufficiently tuned?
Comparing with the existing methods is almost a “compulsory routine” for traffic prediction
studies. Obviously, the proposed method will be carefully and elaborately tuned in order to
achieve the best performance. However, one natural question arises: are the competitive
methods well tuned according to the study scenario? I guess the answer is negative, since
those competitive methods are not within the study scope and, even, “the worse, the better”
for the performance of the competitive methods.
September 24, 2019
5. Is the result significant?
It is common to see that the proposed “well-tuned” and “advanced” method only improve
the prediction accuracy by several percentage points, compared with the traditional methods.
For example, the network-wide prediction error (e.g., MAPE) of a proposed deep learning-
based method is 22%, while the errors of the support vector machine-based method and
historical average method are 25% and 30%, respectively. Therefore, the questions are (1) is
the accuracy like 22% satisfactory? (2) can we say the improvement (about 5% or even less)
is significant? (3) doesn’t the improvement happen occasionally (after all, most of the
proposed methods are only tested in one or two scenarios).
6. Is the newly proposed model reproducible?
Almost no transportation journal compels authors to upload their codes or data and the
authors in the field have no “habit” to opensource their codes as well. Surely, we trust the
authors, but it undoubtedly makes reproducing the algorithm difficult.