A Taxonomy of Traffic Forecasting Regression Problems From a Supervised Learning Perspective

Abstract

One contemporary policy to deal with traffic congestion is the design and implementation of forecasting methods that allow users to plan ahead of time and decision makers to improve traffic management. Current data availability and growing computational capacities have increased the use of machine learning (ML) to address traffic prediction, which is mostly modeled as a supervised regression problem. Although some studies have presented taxonomies to sort the literature in this field, they are mostly oriented to classify the ML methods applied and a little effort has been directed to categorize the traffic forecasting problems approached by them. As far as we know, there is no comprehensive taxonomy that classifies these problems from the point of view of both traffic and ML. In this paper, we propose a taxonomy to categorize the aforementioned problems from both traffic and a supervised regression learning perspective. The taxonomy aims at unifying and consolidating categorization criteria related to traffic and it introduces new criteria to classify the problems in terms of how they are modeled from a supervised regression approach. The traffic forecasting literature, from 2000 to 2019, is categorized using this taxonomy to illustrate its descriptive power. From this categorization, different remarks are discussed regarding the current gaps and trends in the addressed traffic forecasting area.

A graph CNN-LSTM neural network for short and long-term traffic forecasting based on trajectory data

“…TF can be tackled using different modelling perspectives. The four most common approaches found in transportation literature are: (i) the statistical time-series perspective (Ermagun and Levinson, 2018), (ii) the supervised regression problem (Angarita-Zapata et al, 2019;Howell, 2018), (iii) the supervised classification problem (Angarita-Zapata et al, 2018;Lopez-Garcia et al, 2016), and (iv) a clustering-pattern recognition approach (Aldhyani and Joshi, 2018). They are described as follows.…”

“…This paper is focused on the supervised regression approach. The literature on TF reports a wide variety of ML methods for supervised traffic prediction such as Neural Networks, SVM or Random Forest (Angarita-Zapata et al, 2019;Lana et al, 2018). These ML methods have shown satisfactory results when the complexity of the problem is moderate.…”

“…Despite their success, classical ML methods present limitations when capturing the spatial and temporal relationships of traffic patterns (Yang, 2013;Zhao et al, 2018;Zhang et al, 2016). In recent years, Deep Neural Networks (DNNs) have caught the attention of transportation research because of their high representational power that facilitates dealing with the aforementioned components of traffic and their good performance in terms of accuracy and error metrics (Angarita-Zapata et al, 2019;Do et al, 2018;Ermagun and Levinson, 2018).…”

“…However, as far as we know, previous studies on hybrid Graph CNN-RNN architectures were designed to work with point-wise traffic sensors or a reduced set of road links for limiting the complexity and dimension of the input data. Furthermore, they are only tested for short-term time horizons, which correspond to less than 60-minutes-ahead predictions according to the categorisations proposed in recent literature (Lana et al, 2018;Angarita-Zapata et al, 2019). Regarding the former issue, when considering the whole road network it is important to exploit the strengths of DNNs in order to deal with the particularities of trajectory data, since, in this case, traffic information is spread around the whole city and the temporal dependencies between road links are not explicitly represented on the data.…”

Traffic Light Cycle Configuration of Single Intersection Based on Modified Q-Learning

“…Machine learning can be divided into three categories: Supervised learning [21], unsupervised learning [22] and reinforcement learning. The supervised learning algorithm is based on the budget to access the desired output of the limited input (training tag), and optimizes the selection of the input it receives for the training tag.…”