A survey of evaluation methods for personal route and destination prediction from mobility traces

Personal mobility data can nowadays be easily collected by personal mobile phones and used for analytical modeling. To assist in such an analysis, a variety of computational approaches have been developed. The goal is to extract mobility patterns in order to provide traveling assistance, information, recommendations or on-demand services. While various computational techniques are being developed, research literature on destination and route prediction lacks consistency in evaluation methods for such approaches. This study presents a review and categorization of evaluation criteria and terminology used in assessing the performance of such methods. The review is complemented by experimental analysis of selected evaluation criteria, to highlight the nuances existing between the evaluation measures. The experimental study uses previously unpublished mobility data of 15 users collected over a period of 6 months in Helsinki metropolitan area in Finland. The article is primarily intended for researchers developing approaches for personalized mobility analysis, as well as a guideline for practitioners to select criteria when assessing and selecting between computational approaches. Our main recommendation is to consider user-specific accuracy measures in addition to averaged aggregates, as well as to take into consideration that for many users accuracy does not saturate fast and the performance keeps evolving over time. Therefore, we recommend using time-weighted measures. (c) 2017 Wiley Periodicals, Inc.Peer reviewe