On Making More Efficient Location Prediction

There is an increasing interest in mining individuals’ location history for location prediction purpose given the prevalence of location-aware devices. Longer location history generally results in more accurate predictions. Yet, processing long location history for a large number of people can also render poor computational efficiency. Using shorter location history raises the concern of loss of information and thus less accurate predictions. The authors, however, hypothesized that a short location history may be enough to achieve a satisfactory level of prediction accuracy for those individuals whose trajectories contain little uncertainty. The primary objective of this paper is to examine the correlation between the length of location history and prediction accuracy for subpopulations differing in the level of uncertainty in their trajectories. Analyses are performed with a mobile phone data set consisting of the traces of 3,568 individuals over two months. The amount of uncertainty in trajectory is found to be an instrumental indicator of the amount of input information required for location prediction. Specifically, given 100 historical locations, an accuracy level marginally over 80% can be achieved for people with a low level of uncertainty. In contrast, for those with a high level of uncertainty, prediction accuracy can hardly reach 50% with 100 historical locations. This finding allows for customizing the amount of information input in location prediction for subpopulations differing in the amount of uncertainty in trajectory. Being able to discard some data without compromising model prediction accuracy is one way to deal with an overwhelming amount of data.