Active Learning for transport studies: the case of rare or sparse demand samples

Abstract

For two decades now, the rapid rise and wide diffusion of new technologies have enabled the generation of a massive amount data – commonly known as Big-Data. Big-Data have been widely adopted in nearly every field of transportation, from behavioural analysis, to traffic predictions, or model calibration. While the initial promise of Big-Data was to allow for a better understanding of the total population and their diversity, years of research prove that this is not always the case. The main limitation is that Big-Data are difficult to interpret. For instance, if we use mobile phone network data to create a demand matrix, the model will have a bias (penetration rate, service provider, multiple devices) which is difficult to measure.

In general, the data should be representative of the entire population and represent all different demand segments in a correct way. This is currently not the case for many so-called big-data sources and, anyway, it is not easy to assess the representativeness of a given data-set. This paper proposes using active learning to address this problem. Active learning models use machine learning algorithms to sample data from a population (or dataset) and create a small yet representative set of observations that encompass the main attributes of the entire population. To that end, we introduce an enhanced active learning algorithm that combines two models. Traditional Active Learning Techniques (such as Gaussian Processes) are used to sample supply-related data. A heuristic model based on the Branch and Bound algorithm is instead used to sample demandrelated information. By combining the two models, the proposed approach can sample supply and demand data together. The model is used to sample origin-destination trips for bike-sharing from sparse demand matrices. The case study uses real world data from New York city, showing promising results.