Bus line trajectories classification using weightless neural networks

Abstract

Geo-enabled devices are ubiquitous nowadays. Within a diversity of possible applications using the huge of amount data generated by this technology, our work focuses on a chronic problem of Rio de Janeiro city: its public bus system. This text presents a framework for GPS trajectories classification, whose focus is the identification of bus routes of a public bus system. In order to do that, it was used the lightweight and versatile WiSARD, a weightless neural network classifier. Different binarization methods were used to adapt raw data to WiSARD’s binary input, making use of a set of rules defined by the application domain. Yet, it is evaluated a way of combining WiSARD through decision directed acyclic graphs. All these approachs result in different flavors of a neuro-symbolic learning system. The framework was tested against a vast data set created from open access and real-time data acquired from the current bus system of Rio de Janeiro city. Results obtained suggest the applicability of the proposed solution in a classification problem with more than 500 classes. Comparisons made also indicate an equivalent performance of WiSARD and other state-of-art and widely used machine learning methods. In addition, the framework described here is believed to be adaptable to other application domains.