Inferring the Activities of Smartphone Users from Context Measurements Using Bayesian Inference and Random Utility Models
R Hurtubia, G Flötteröd, M Bierlaire, EPFL - Transport and Mobility Laboratory, CH
The increasing availability of new technologies such as smart-phones and GPS devices provides a wealth of individual-level mobility information, which can be used to forecast the user's future activities as well as the locations chosen to perform these activities. These predictions can be used to provide information that can help the user to make better decisions. For example, customized information about activity opportunities and traffic conditions can be delivered to the user depending on predicted activities and locations.
In this paper we propose a Bayesian tracking approach to this problem, combining prior knowledge, given through a random utility model of the user's short-term activity and location choices, with ambient real-time information (measurements) gathered by a smart-phone.
More specifically, the discrete choice model provides, at every point in discrete time, a prior distribution of the individual?s upcoming activities and activity locations. The explanatory variables of this model are the individual's memory (possibly aggregated representations of the individual's past activities and locations), individual-specific parameters such as its socioeconomics, and land use characteristics. Combining this model with one of the individual's information processing (memory update), we obtain a dynamical transition equation of the individual's ?state? which consists of its current location, activity, and memory. At any point in time, this transition equation provides a prior distribution of the individual's future states.
For the discrete choice model we propose a conditional probability structure, where future location choice depends on the activity to perform and, at the same time, future activities depend on the ?history? of the individual?s previous states. However this choice structure is not definitive and other specifications are tested. Given this model, we are able to generate a probabilistic offline prediction of what we expect the considered person to do throughout the upcoming day.
The spectrum of useful sensor data provided by a smart-phone ranges from GPS signals to ambient visual and acoustic information. However, our first experiments exclusively account for GPS measurements. Further ambient information can consistently be incorporated in the Bayesian tracking framework described next.
In a Bayesian setting, our knowledge about the current state of a person is expressed in terms of a probability assigned to every possible state. From this, future states can be predicted in a probabilistic manner by recursive application of the state transition model. Upon arrival of new sensor data, we update our current state estimate by multiplying the previously predicted probability of each state by the likelihood of the new measurements given the state. That is, we ?modulate? the probabilities such that states that are consistent with the measurements receive an increased probability. Recalling that the notion of a ?state? comprises the location and activity of an individual, this solves the activity/location tracking and prediction problem in a probabilistic setting. By explicitly accounting for the distribution of possible results, we are also able to make statements about the reliability of our estimates.
The random utility model is estimated over a database containing land-use information for the city of Lausanne and the Swiss Transport Microcensus. The spatial disaggregation of the land-use database defines the zoning for our model, which consists of 100x100 meters gridcells. The transport microcensus includes an activity survey which indicates performed activities during the day and their corresponding location for a group of individuals in Lausanne. The real time measurements come from smart-phones provided to a group of individual in the city of Lausanne who participate on a survey to track their movements during the day and the activities performed in each location. For first experimental investigations, real time measurements are simulated from the information contained in the microcensus. We experimentally evaluate the feasibility of the proposed approach and identify further ambient information that is relevant for the improvement of the estimation precision.
Association for European Transport