Multi-label classification of estimated time of arrival with ensemble neural networks in bus transportation network

Abstract

The arrival times of buses are often hard to predict due to variation of real time traffic conditions, deployment schedules and traffic incidents. The provision of timely arrival time information is thus vital for passengers to minimize their waiting time and improve riders' confidence in the public transportation system, directly promoting more ridership. Multiple buses are commonly observed to arrive at a bus stop every hour. In this research, the prediction of estimated time of arrival (ETA) of buses is translated into a multi-label classification problem. Using buses historical global positioning system (GPS) arrival time, neural network models (ANN) are shown to be reliable solutions for the problem, and ensemble of neural networks are explored for more relevant output. The experimental results demonstrate that 77–78% of the time, ANN models are able to accurately predict arrival time of buses. The neural network models are able to outperform the other algorithms (i.e. Decision Tree, Random Forest, Naive Bayes) in the classification of multi-label arrival times by up to 8% based on performance metrics such as hamming loss, accuracy, and F 1 scores. Among the ensemble models, Random A-labelsets (RAkEL) with 2-labelsets and 3-labelsets outperform the single neural network by 2–3% in multi-label accuracy and F1 scores. The findings are convincing for us to further explore other problem transformation and ensemble approaches in the future.