EASTAR: Event-Anchored and Task-Adaptive Urban Region Representation Learning

EASTAR anchors mobility semantics with activity events and dynamically routes channels per task, reducing MAE by up to 39.15%.
EASTAR is an event-anchored and task-adaptive urban region representation learning framework that addresses semantic entanglement in aggregated OD matrices. It uses offline activity events as sparse semantic anchors to decompose mobility representations into purpose-related semantic channels, with a free channel capturing uncovered patterns and a global channel preserving spatial structure. A task-adaptive prompt router initializes task-channel correlations from LLM priors and refines them via downstream supervision, enabling dynamic semantic selection. On four prediction tasks across two real-world urban datasets, EASTAR consistently outperforms baselines, reducing MAE by up to 39.15%.
Urban region representation learning is a core task in urban computing, aiming to encode regional functions and interactions into low-dimensional vectors to support various urban analyses. Human mobility data has become a key signal because it directly reflects inter-region interactions. However, most existing methods learn representations based on aggregated Origin-Destination (OD) matrices, where trips with different activity purposes are mixed into "purpose-agnostic" interaction patterns. Although static attributes such as POI and land use can enrich regional semantics, they cannot explicitly anchor mobility interactions to activity purposes. This causes downstream tasks to adapt only on entangled embeddings, making it difficult to select task-relevant functional semantics.

To address these issues, researchers propose EASTAR—an event-anchored and task-adaptive urban region representation learning framework. Its core idea is to use offline activity events as sparse semantic anchors, decomposing aggregated mobility representations into multiple purpose-related semantic channels; simultaneously, a free channel is introduced to capture latent mobility patterns not covered by event data, and a global channel is set to preserve overall spatial structure. To adapt to different downstream objectives, EASTAR designs a task-adaptive prompt router that initializes task-channel correlations from LLM-derived priors and refines routing weights through downstream supervision. Across four prediction tasks on two real-world urban datasets, EASTAR consistently outperforms representative baselines, reducing MAE by up to 39.15%. Ablation and interpretability analyses further validate the effectiveness of event anchoring, semantic channel construction, and task-adaptive routing.
Problem Background: Purpose-Agnostic Mobility Patterns Struggle to Support Downstream Tasks
Human mobility data is the most direct interaction signal in region representation learning, but aggregated OD matrices mix trips with different purposes—commuting, shopping, leisure—into purpose-agnostic interaction patterns. This mixing makes learned embeddings hard to distinguish functional semantics, and downstream tasks (e.g., region function prediction, flow forecasting) can only adapt within an entangled representation space, unable to select relevant semantics for specific tasks. While static urban attributes (POI, land use) can supplement region-level semantics, they lack explicit purpose anchoring with mobility interactions and cannot resolve semantic entanglement.
OD matrix (Origin-Destination Matrix) is a fundamental data structure in urban mobility analysis, where rows and columns correspond to trip origin and destination regions, and cell values represent trip flows between two regions. OD matrices aggregated from taxi trajectories, mobile phone signaling, or card swipe records can characterize the overall pattern of urban spatial interactions and are common inputs for tasks such as traffic forecasting and region function identification. However, the aggregation process discards trip purposes, temporal context, and individual attributes, retaining only flow counts. Consequently, heterogeneous trips such as commuting, shopping, and medical visits are compressed into the same numerical signal, and learned region embeddings struggle to distinguish functional differences. To recover purpose semantics, external activity event data (e.g., concerts, exhibitions, sports events) is typically introduced as anchors to associate trip flows at specific spatiotemporal points with explicit activity types, thereby providing supervision signals to disentangle mobility patterns.
EASTAR Framework: Decomposing Semantic Channels via Event Anchoring
EASTAR treats offline activity events as sparse semantic anchors, using them to decompose aggregated mobility representations into multiple purpose-related semantic channels. Each channel corresponds to a category of activity purpose (e.g., work, consumption, entertainment), explicitly anchoring mobility interactions to purpose signals. Additionally, the framework introduces a free channel to capture latent mobility patterns not covered by event data, avoiding information loss due to event sparsity; meanwhile, a global channel preserves overall spatial structure, ensuring the completeness of region representations. This multi-channel design achieves disentanglement of functional semantics while maintaining spatial structure.

Semantic channel borrows the "channel" concept from multi-task learning and disentangled representation learning: splitting multiple semantic factors originally entangled in a single embedding space into several independent subspaces for separate modeling. In EASTAR, each channel corresponds to a category of activity purpose, and representations within a channel encode only mobility interaction patterns related to that purpose, with information redundancy between channels suppressed as much as possible. This decomposition allows downstream tasks to select channels on demand rather than passively adapting on mixed representations. The accompanying free channel absorbs residual mobility patterns not covered by event anchors, acting as a fallback to prevent information omission due to sparse events; the global channel preserves overall spatial structure without distinguishing purposes, maintaining basic inter-region accessibility and adjacency relationships. The three types of channels complement each other, jointly forming a region representation that is both disentangled and complete.
Task-Adaptive Routing: From LLM Priors to Downstream Supervision
To adapt the learned channel space to different downstream objectives, EASTAR designs a task-adaptive prompt router. The router first initializes task-channel correlations using LLM-derived priors, then refines routing weights through supervision signals from downstream tasks. This means different tasks can dynamically select their most relevant semantic channels rather than performing uniform adaptation on fixed embeddings. This design aligns semantic selection with task objectives, improving the transferability and task specificity of representations.
The "prompt" in the prompt router borrows the concept of prompts from the large language model field: task descriptions (e.g., "predict regional foot traffic," "infer regional function") are converted into text prompts, and the LLM outputs initial relevance scores between the task and each semantic channel. The role of this prior is that when downstream labeled data is scarce, routing weights are not randomly initialized but fine-tuned from a semantically reasonable starting point. Subsequently, supervision signals from downstream tasks refine the routing weights through gradient backpropagation, gradually aligning task-channel correspondences with true prediction objectives. This two-stage design of "LLM prior + supervised fine-tuning" essentially transfers the adapter or prompt tuning ideas from parameter-efficient transfer learning to the urban representation learning scenario, enabling the same set of channel representations to adapt to multiple heterogeneous tasks at low cost.
Experimental Validation: MAE Reduced by Up to 39.15%
On two real-world urban datasets, EASTAR consistently outperforms representative baselines across four prediction tasks, reducing MAE by up to 39.15%. Ablation experiments and interpretability analyses further demonstrate: event anchoring effectively improves the purpose discriminability of semantic channels; semantic channel construction (including free and global channels) significantly contributes to performance; task-adaptive routing can adjust channel weights according to task objectives, and routing results are interpretable. These results validate the effectiveness of anchoring mobility interactions to activity purposes and dynamically selecting semantic channels per task.
Significance and Outlook
The core contribution of EASTAR lies in introducing "event anchoring" and "task adaptation" into urban region representation learning, solving the problem of purpose semantics loss in aggregated mobility data. Through semantic channel decomposition and LLM-prior-guided routing, the framework can provide interpretable, task-relevant region representations for different downstream tasks. In the future, this method is expected to extend to more urban analysis scenarios and further explore handling event data sparsity and cross-city transferability.
One-line Summary
EASTAR anchors mobility semantics with activity events and dynamically routes channels per task, reducing MAE by up to 39.15%.
Paragraph Summary
EASTAR is an event-anchored and task-adaptive urban region representation learning framework designed to solve the semantic entanglement caused by mixing trips with different purposes in aggregated OD matrices. It uses offline activity events as sparse semantic anchors to decompose mobility representations into multiple purpose-related semantic channels, supplemented by a free channel to capture uncovered patterns and a global channel to preserve spatial structure. The framework also designs a task-adaptive prompt router that initializes task-channel correlations from LLM priors and refines weights through downstream supervision, enabling different tasks to dynamically select relevant semantics. On four prediction tasks across two real-world urban datasets, EASTAR consistently outperforms representative baselines, reducing MAE by up to 39.15%, with ablation and interpretability analyses validating the effectiveness of each module.
Related articles

Gemini 4 Neon Released: A Balance Between Frontier Intelligence and Efficient Task Execution
Google DeepMind releases Gemini 4 Neon, focused on frontier intelligence and fast end-to-end task execution. This article breaks down its benchmark performance in coding, enterprise analysis, scientific reasoning, and cybersecurity, its four thinking levels, pricing, and safety alignment data.

Is AI the Culprit Behind Global Outages? The Trust Crisis Behind the NSA and FAA Failures
Fourth major failure in five days hits NSA classified network. Microsoft blames a software bug, FAA denies its new AI system, NSA says routine maintenance. Senator Warner suggests AI accelerates vulnerability discovery, leading to rushed fixes — while 39% of Americans already view AI negatively, exposing a deepening trust crisis.

Robot Developer Builds Custom B-rep Kernel with 3,487 Tests
A robot developer builds a custom B-rep geometry kernel with 3,487 tests, exploring why robotics needs dedicated CAD tools and the barriers to building one.