Adaptive Streaming PPM for Human Mobility Prediction under Memory Constraints - Synthetic Dataset

The synthetic dataset is procedurally generated to stress-test sequential modeling capabilities under controlled, yet realistic, human mobility conditions. Spanning an extensive 400-day simulation period, the generative algorithm is designed to produce trajectories that evaluate the limits of both fixed-order Markov and variable-order models. The data generation pipeline explicitly injects the following behaviors: Periodic Commuting Structures: The algorithm defines distinct routines based on the day of the week, alternating between standard workplace commuting on weekdays and stochastic exploratory behavior on weekends. Long-Range Temporal Dependencies: The simulation enforces deep sequential dependencies where evening locations depend heavily on morning choices (e.g., visiting a gym in the morning strictly prevents visiting the gym in the evening). GPS Noise & Spatial Bouncing: To simulate sensor inaccuracy, random Gaussian noise (0, 0.0005) is injected into the raw latitude and longitude coordinates, forcing the spatial clustering layer to dynamically map noisy GPS points into stable H3 hexagons. Stochastic Transitions: Interspersed within the daily routines are probabilistic detours to secondary locations (cafes, supermarkets, parks) to prevent the trajectories from becoming trivially deterministic. Abrupt Behavioral Shifts (Concept Drift): To explicitly evaluate memory bounding and model adaptation, a sudden “workplace change” is injected precisely at day 100. The agent permanently abandons the historical workplace coordinate and adopts a new spatial destination, acting as a critical non-stationary concept drift validation test.