Constructing multi-labelled decision trees for junction design using the predicted probabilities

In this paper, we evaluate the use of traditional decision tree algorithms CRT, CHAID and QUEST to determine a decision tree which can be used to predict a set of (Pareto optimal) junction design alternatives (e.g. signal or roundabout) for a given traffic demand pattern and available space. This is a multi-label decision tree problem. Traditional decision tree algorithms can normally not deal with multiple target labels, since they aim to produce trees with single target labels. However, we propose an approach in which we normalise the training data and use the predicted probabilities of the resulting tree, confronted with a threshold value, to determine multiple target labels. This enables us to predict sets of junction design alternatives with traditional algorithms and thus having the advantage of using profoundly proven and widely available methods with a range of modelling options. We evaluate our approach based on its performance concerning tree complexity and predictive accuracy, for which we introduce new set comparison measures. We test our approach with different experimental runs varying the algorithms, parameters and threshold values. The results show that it is possible to determine decision trees which can be used to predict sets of junction design alternatives with 82-90% accuracy.