Evaluating Automated Machine Learning on Supervised Regression Traffic Forecasting Problems

Abstract

Traffic Forecasting is a well-known strategy that supports road users and decision-makers to plan their movements on the roads and to improve the management of traffic, respectively. Current data availability and growing computational capacities have increased the use of Machine Learning methods to tackle Traffic Forecasting, which is mostly modelled as a supervised regression problem. Despite the broad range of Machine Learning algorithms, there are no baselines to determine what are the most suitable methods and their hyper-parameters configurations to approach the different Traffic Forecasting regression problems reported in the literature. In Machine Learning, this is known as the Model Selection Problem, and although Automated Machine Learning methods has proved successful dealing with this problem in other areas, it has hardly been explored in Traffic Forecasting. In this work, we go deeply into the benefits of Automated Machine Learning in the aforementioned field. To this end, we use Auto-WEKA, a well-known AutoML method, on a subset of families of Traffic Forecasting regression problems characterised by having loop detectors, as traffic data source, and scales of predictions focused on the point and the road segment levels within freeway and urban environments. The experiments include data from the Caltrans Performance Measurement System and the Madrid City Council. The results show that AutoML methods can provide competitive results for TF with low human intervention.

A Case Study of AutoML for Supervised Crash Severity Prediction

“…There is no Au-toML method that consistently outperforms all Au-toML competitors. The latter is consistent with other results of AutoML in transportation problems wherein longer run times do not necessarily lead to drastically better results [6,7]. For the case of As (based on metalearning, optimization, and ensemble learning) and Ag (based bagging, stacking, and ensemble learning), they get slight score improvements as their allocated time budget increases.…”

“…In the transportation area, to the best authors' knowledge, only a few papers have used AutoML methods in this knowledge domain [24,8,6]. The first research carried out by Vlahogianni et al [24] proposed a surrogate modeling technique that optimizes both the algorithm selection and the hyperparameter setting.…”

“…After that, Angarita et al in [7,6,8] tested Auto-Sklearn and AutoWEKA [23] in the prediction of traffic modeled as a supervised learning problem. Those papers compared the performance of AutoML w.r.t.…”

“…AutoML methods have been successfully used in other areas; however, an extensive analysis to determine their strengths and weaknesses has not been carried out in very diverse learning tasks, such as crash severity forecasting. To the best of our knowledge, only a few papers have applied concepts of AutoML in the transportation domain, specifically in supervised traffic forecasting [6,7,8,24]. Thus, other transportation areas, such as crash severity prediction that poses multiple challenges to the ML paradigm, remains unexplored by AutoML.…”

A Bibliometric Analysis and Benchmark of Machine Learning and AutoML in Crash Severity Prediction: The Case Study of Three Colombian Cities

“…AutoML aims at automatically finding competitive ML pipelines (the combination of preprocessing techniques and an ML algorithm) that maximize a performance metric on given data without being specialized in the problem domain from which the data are derived (generalpurpose AutoML) [9]. AutoML methods have been successfully used in other areas of the transportation domain, like traffic forecasting [10][11][12][13]. Nevertheless, the extent to which general-purpose AutoML can be competitive in diverse transportation problems, such as CSP, is far from being fully answered.…”