Weakly Supervised Framework for Trip Purpose Inference from GPS Data
A novel weakly supervised approach has been introduced for deducing trip intentions from GPS data. This technique combines semantic zones of points of interest (POIs) at the neighborhood level with distance-weighted spatial probabilities, employs distinct inference methods for essential and non-essential activities, and utilizes a multi-phase Pareto optimization to decrease distributional divergence from household travel survey data, all without the need for labeled annotations. Tested on more than 81 million staypoints in Los Angeles, this framework achieves a 23% reduction in activity type frequency Jensen-Shannon distance (JSD), a 48% decrease in start time JSD, and a 12% drop in duration JSD.
Key facts
- Framework uses weakly supervised learning for trip purpose inference.
- Integrates POI semantic zones and distance-weighted spatial likelihoods.
- Differentiates inference strategies for mandatory and non-mandatory activities.
- Multi-phase Pareto optimization minimizes distributional divergence from survey statistics.
- Evaluated on over 81 million staypoints in Los Angeles.
- Reduces activity type frequency JSD by 23%.
- Reduces start time JSD by 48%.
- Reduces duration JSD by 12%.
Entities
Locations
- Los Angeles