Lifelong Synthetic Data Stream with Concept Drift

The Lifelong Synthetic Stream with Concept Drift is a custom-built dataset designed to stress-test the structural elasticity and memory bounds of the models over an extended temporal scale. It simulates a 3,000-day continuous trajectory for a single user. To evaluate resilience against the ”state explosion” problem, the generator injects strict Concept Drifts by relocating the user to a completely new simulated city (a distinct geographic coordinate base) every 180 days. Furthermore, the daily routines are explicitly designed to contain high-order sequential bottlenecks. For example, morning sequences like Home → Metro → Work diverge from evening sequences like Work → Metro → Gym. In these cases, the intermediate Metro state requires Order-2 awareness to resolve correctly. Finally, a 20% exploration noise factor introduces random, non-recurring Points of Interest (POIs) to artificially inflate the transition space. This aggressively tests the Space-Saving eviction mechanism under high-cardinality conditions.