Nextact: A Hybrid Approach For High-Resolution Human Activity Predictions

Emerging anticipatory applications that intervene or proactively support users in daily life require an in-depth understanding of prospective human behavior. However, accurate and fine-grained prediction of human behavior is complex and conventional machine learning approaches show limitations. In this paper, we present a novel approach, namely NextAct, to predict human activities by utilizing temporal and spatial features predicted from calendar events, user routines and sequences of next place predictions for remaining uncertain time slots in between. NextAct hereby makes spatial features in addition to the obvious temporal features available for the human activity predictor. We evaluate our hybrid approach on a four-week user-annotated dataset from 30 participants; we compare our forecasting results against prediction techniques proposed in the literature. The results show an F-1-score up to 82.6% and a performance gain up to 16.6% compared to state of the art approaches. We also discuss the cold-start problem of individual models and show that we achieve adequate results after 4 days. NextAct opens a novel way for human activity prediction to proactively support a user and optimize his next activity or day.