Skip to main content
Back to timeline
arXivSource publication:

Urban trip matrices stop mixing purposes: EASTAR splits mobility into event-anchored semantic channels, cutting Chicago check-in MAE by 39.15%

Synopsis

Urban region representations have long relied on aggregated origin-destination matrices, where commuting, shopping and social trips are mixed into one purpose-agnostic interaction strength, making task-relevant functional semantics hard to select. EASTAR treats offline activity events as sparse semantic anchors, completes them along mobility flows, and factorizes the aggregated mobility representation into purpose-related channels while keeping free channels and a global channel. On two cities and four prediction tasks, EASTAR outperforms the representative baselines it is compared with, including a 39.15% MAE reduction on Chicago check-in prediction relative to the strongest baseline.

Source-provided article image: EASTAR: Event-Anchored Semantic and Task-Adaptive Region Representation Learning
arXiv

Interpretation

Urban region representations can now be split by activity purpose instead of keeping a single purpose-agnostic mobility interaction map. Earlier mobility-based methods learn from aggregated origin-destination matrices, mixing trips driven by different activity purposes into purpose-agnostic interaction patterns; static attributes such as POIs and land use describe what facilities a region contains but do not ground interactions in activity purpose. On two real urban datasets, New York City and Chicago, across four prediction tasks, EASTAR outperforms 12 representative baselines, reducing MAE by 39.15% on Chicago check-in prediction and 23.11% on New York check-in prediction relative to the strongest baselines.

The split is grounded in offline activity events rather than an added data view. Event data covers only part of the regions and records activity category, time, location and participant counts, so it is sparse. Each region's events are aggregated by category into text documents and encoded into semantic anchors by a sentence transformer, then propagated over the mobility graph with flow-aware attention to fill missing categories; the completed semantics serve as queries and the mobility representation as keys and values in broadcast attention, extracting one channel per purpose, with free channels for latent patterns events miss and a global channel for overall spatial structure.

Once built, channels are selected per task rather than shared with one weight set. Tasks such as crime and population depend on different functional semantics, so a single task-agnostic embedding cannot fit all objectives equally well. A large language model turns the task description and channel semantics into an initial channel-relevance prior that initializes a lightweight router, which is then refined under downstream supervision; across New York and Chicago, different tasks activate different channel combinations, with check-in prediction favoring Food and Social channels and crime prediction relying on the Global channel supplemented by Social dynamics.

Perspective

The demonstrated scope is two cities, New York and Chicago, and four region-level count regression tasks, with 180 and 77 non-overlapping regions and event semantics drawn from Meetup activity records.

Event data serve as sparse semantic anchors rather than ground-truth trip-purpose labels, and evaluation covers only two cities; broader urban contexts and richer activity sources remain future work.