A Collaborative Filtering Approach to Citywide Human Mobility Completion from Sparse Call Records / 2500
Zipei Fan, Ayumi Arai, Xuan Song, Apichon Witayangkurn, Hiroshi Kanasugi, Ryosuke Shibasaki
Most of human mobility big datasets available by now, for example call detail records or twitter data with geotag, are always sparse and heavily biased. As a result, using such kind of data to directly represent real-world human mobility is unreliable and problematic. However, difficult though it is, a completion of human mobility turns out to be a promising way to minimize the issues of sparsity and bias. In this paper, we model the completion problem as a recommender system and therefore solve this problem in a collaborative filtering (CF) framework. We propose a spatio-temporal CF that simultaneously infers the topic distribution over users, time-of-days, days as well as locations, and then use the topic distributions to estimate a posterior over locations and infer the optimal location sequence in a Hidden Markov Model considering the spatio-temporal continuity. We apply and evaluate our algorithm using a real-world Call Detail Records dataset from Bangladesh and gives an application on Dynamic Census, which incorporates the survey data from cell phone users to generate an hourly population distribution with attributes.