Activity Inference through Sequence Alignment

Activity inference attempts to identify what a person is doing at a given point in time from a series of observations. Since the 1980s, the task has developed into a fruitful research field and is now considered a key step in the design of many human-centred systems. For activity inference, wearable and mobile devices are unique opportunities to sense a user's context unobtrusively throughout the day. Unfortunately, the limited battery life of these platforms does not always allow continuous activity logging. In this paper, we present a novel technique to fill in gaps in activity logs by exploiting both short- and long-range dependencies in human behaviour. Inference is performed by sequence alignment using scoring parameters learnt from training data in a probabilistic framework. Experiments on the Reality Mining dataset show significant improvements over baseline results even with reduced training and long gaps in data.