Source-linked AI summary
Spatial-Knowledge-Graph-Grounded LLM Agents for Neighborhood Livability Evaluation
Haiyan Hao
TL;DR
Static accessibility indicators do not fully capture how mobility capacities, schedules, and household responsibilities shape neighborhood experience. This paper develops a spatial-KG-grounded LLM-agent prototype that generates and revises household schedules, checks and routes them, and supports traceable resident interviews. The demonstration shows that facility availability does not ensure convenient access, while simulated interpretations remain distinct from observed perceptions.
Problem
Static indicators describe spatial opportunities but rarely represent feasible activity chains, household coordination, or resident-specific livability experiences.
Method
A spatial knowledge graph grounds household-level LLM schedule generation and revision, rule-based feasibility review, network materialization, and evidence-grounded resident interviews.
Results
Facility availability did not ensure convenient access: mobility-sensitive residents had no trips within 15 minutes despite reaching required destinations, and child-related trips required accompaniment.
Takeaways & Limitations
The prototype connects spatial opportunity, household activity constraints, and resident-specific livability interpretation through an auditable simulation framework.
Takeaways & Limitations
Interview outputs interpret a modeled environment rather than providing direct evidence of residents’ perceptions or policy preferences.
Abstract
from arXiv · showhide
Neighborhood livability is commonly assessed with static built-environment indicators, such as facility proximity, street connectivity, and access to public space. These measures describe available opportunities but do not directly represent how residents with different mobility capacities, household roles, schedules, and care responsibilities experience the neighborhood. This paper presents a prototype framework that uses a spatial knowledge graph (KG) and large language models (LLMs) to generate and revise household schedules, followed by rule-based feasibility checking and GIS-based network materialization. The spatial KG integrates residents, residences, facilities, neighborhood context, and sampled road hubs; Graph-RAG retrieves each household's nearby spatial context, including candidate POIs and approximate walking times, for the scheduling LLM. The LLM produces structured household schedules, while rules are used for lightweight repairs and auditable feasibility checks. The LLM then revises schedules in response to identified feasibility issues. A routing module derives the actual travel paths, travel times, modes, and event histories from the road network. The resulting events support synthetic resident-agent interviews about daily convenience, travel burden, activity feasibility, and household coordination. A prototype demonstration in a Shenzhen neighborhood shows that nominal facility availability does not necessarily imply convenient access: residents with limited mobility and households with care responsibilities experience greater travel and coordination burdens. The framework offers an auditable way to connect spatial opportunity, household activity constraints, and resident-specific livability interpretation, while keeping simulated experience distinct from observed perception.
1. Introduction
Conventional livability indicators describe neighborhood opportunities but do not capture how residents with different capacities, schedules, and household responsibilities experience them. The paper addresses this gap with a spatial-KG-grounded LLM-agent framework for simulating household schedules and producing traceable resident-specific evaluations.
- Static indicators measure services, public space, transport, walkability, and safety, but mainly describe neighborhood provision rather than residents’ daily experiences.
- The same environment can impose different burdens because mobility capacity and household coordination affect whether nearby facilities are practically reachable.Examples include older residents with slow walking speeds and households coordinating work, school pick-up, childcare, and shopping.
- LLM agents can represent heterogeneous personas and generate plausible schedules, but ungrounded outputs may conflict with actual destinations, routes, and travel conditions.
- The proposed framework simulates household activity-mobility schedules and uses resulting event histories to support evidence-grounded, post-simulation resident interviews.
- Its contributions connect personalized activity feasibility, household circumstances, spatial-KG context, and traceable resident-specific interpretation within an auditable evaluation framework.
2. Related Work
Related work shows that conventional accessibility measures capture spatial opportunities but incompletely represent feasible activity chains, household coordination, and lived evaluation. LLM agents add flexible behavioral representation, yet spatial grounding and traceable translation from schedules to resident-specific livability assessment remain unresolved.
- Neighborhood Livability and Accessibility: Livability research commonly operationalizes neighborhood conditions through spatial indicators, while these measures rarely represent activity chaining or household coordination under daily constraints.
- Activity-Based and Household-Aware Accessibility: Accessibility depends on destination relationships, opening times, activity sequence, available time, mobility resources, and fixed work or school anchors.
- Activity-Based and Household-Aware Accessibility: Household interactions such as escorting children, accompanying older relatives, and dividing care tasks alter accessibility benefits generated by travel decisions.
- LLM Agents in Urban Simulation: LLM agents can represent preferences, adaptive decisions, and natural-language interaction that are difficult to encode exhaustively in rule-based systems.
- LLM Agents in Urban Simulation: Separating semantic schedule generation from spatial realization can produce behaviorally plausible schedules that become inefficient or infeasible after destinations, routes, and travel times are assigned.
- LLM Agents in Urban Simulation: Spatially grounded LLM agents remain underused for translating household schedules and network-derived consequences into traceable, resident-specific neighborhood assessments.
3. Framework and Prototype System
The prototype links spatial data, household personas, iterative scheduling, network materialization, and evidence-grounded interviews through four connected modules. It uses a spatial KG and routable road graph to ground household schedules while retaining configurable personas and inspectable simulation outputs.
- System Workflow and Spatial Knowledge Graph: The system comprises Spatial Setup and KG Construction, Household and Persona Configuration, Iterative Scheduling and Network Materialization, and Inspection and Interviews.
- System Workflow and Spatial Knowledge Graph: The spatial environment combines a user-defined boundary, routable road graph, and facility set, with spatial data organized into a spatial knowledge graph.
- Iterative Scheduling and Network Materialization: For each household, the system retrieves a household-centered ego-subgraph, generates and reviews schedules through an iterative LLM-rule process, and materializes accepted schedules on the road graph.
- Implementation: A web-based implementation returns KG summaries, schedules, event records, aggregate metrics, and interview responses through React/TypeScript and Python FastAPI components.
- Household and Persona Configuration: Structured personas encode mobility capacity, travel modes, activity needs, behavioral constraints, fixed anchors, temporal commitments, and caregiving relationships.
- Household and Persona Configuration: The prototype supports predefined profiles and custom personas for standardized comparisons or scenario-specific analyses.
Spatial KG Construction and Retrieval
The framework organizes residents, households, residences, facilities, activity categories, and road hubs in a spatial knowledge graph. Graph-RAG retrieves household-centered candidate facilities with approximate walking times to ground schedule generation in bounded spatial context.
- The spatial KG represents neighborhoods, residents, households, residences, facilities, activity categories, and sampled road hubs with spatial and household relationships.
- Facilities are grouped by category, ranked by haversine distance, and retained up to ten per category within 3 km through NEAR relations.
- Each NEAR relation stores straight-line distance and approximate walking time at a reference speed of 70 m/min for retrieval and preliminary planning.
- Graph-RAG retrieves a household-centered subgraph by linking members to their shared residence and then to candidate facilities through NEAR relations.
- The retrieved context supplies household attributes, residence information, valid POI identifiers, facility categories, and approximate walking times, limiting unsupported facility invention.
Household Schedule Generation and Review
The scheduling LLM generates structured household schedules from personas and retrieved spatial context, while deterministic checks repair or diagnose feasibility problems. Unresolved schedules are revised iteratively and recorded as failures if they exceed the review limit.
- The scheduling LLM receives household personas, an ego-subgraph, a simulation horizon, scheduling requirements, and an output schema, then returns a TSV household schedule.
- Rule-based review performs conservative repairs, including duplicate removal, joint-activity merging, school-escort normalization, and adjacent-work-segment joining.
- Feasibility checks enforce valid timing and duration, participant exclusivity, spatiotemporal continuity, mandatory work or school anchors, care constraints, and joint-activity consistency.
- Care constraints require adult accompaniment for infants and children under 12 during applicable activities and prohibit simultaneous escorting and participation in another activity.
- Non-anchor destinations must belong to the retrieved candidate set, and activity types must belong to the supported taxonomy.
- When issues remain, the LLM receives diagnostics and revises the TSV schedule; runs exceeding the maximum review rounds are recorded as scheduling failures.
Network Materialization and Event Construction
Accepted schedules are materialized on the road network into routed, mode-assigned events with participant histories and provenance. These outputs support maps, timelines, weekly mobility summaries, and evidence-grounded resident interviews.
- After acceptance, activities are ordered by day and departure time, then mapped from current origins to destinations using fixed work or school anchors and retrieved POIs.
- Trips are snapped to nearest road-network nodes and routed with NetworkX using distance-weighted shortest paths, while travel modes follow resident availability and estimated walking time.
- Origins inherit from preceding realized events, and a member-level clock delays activities, tracks participant locations, inserts return-home movements, and checks overlaps.
- Each event records participants, timing, activity type, origin, destination, network path, realized mode, distance, travel time, materialization success, and failure reason.
- The workflow aggregates materialized household events into weekly metrics while retaining identifiers and metadata for retrieved contexts, generated decisions, repairs, diagnostics, and network outcomes.
- The inspection interface presents route maps, daily timelines, activity histories, and summary indicators, while selected interview evidence prioritizes failed trips, journeys exceeding 15 minutes, service activities, and longer travel times.
- Interview responses are generated from respondent-specific event evidence, linked to supporting records, with unassessable conditions identified from the available data.
- The questionnaire covers daily activities and mobility, while broader dimensions such as social interaction, community participation, and safety require evidence beyond the prototype.
4. Prototype Demonstration and Evaluation
The Shenzhen prototype simulates heterogeneous households, materializes schedules into mobility events, and uses those traces for resident-specific livability evaluation. Results show that activities were generally feasible, but travel burden, facility distribution, mobility constraints, and household coordination produced meaningful differences across residents.
- Demonstration setup: The demonstration instantiated six households and 11 resident agents within a manually defined Shenzhen neighborhood boundary.Five households came from predefined profiles, and a recently married couple without children was added as H6.
- Demonstration setup: The prototype interface maps neighborhood boundaries, facilities, roads, activity-mobility trajectories, and time-filtered simulation outputs.Displayed outputs include trip counts, completed and failed activities, trips within 15 minutes, and mean travel time.
- Schedule and event outcomes: Aunt Chen completed all 12 out-of-home activities and eight return-home events, but mean travel time was 25.1 minutes and only 10.0% of trips met the 15-minute threshold.Her activity-specific mean travel times ranged from 9.8 minutes for pharmacy visits to 31.3 minutes for park visits.
- Schedule and event outcomes: Across 11 agents, the simulation generated 172 unique event records, with mean travel time ranging from 30.7 minutes for Qiang Wu to 17.9 minutes for Xiao Hong.Aunt Chen had 10.0% of trips within 15 minutes, compared with 36.4% for Mr. Gao and Xiao Hong.
- Household coordination: Children and care-dependent residents required household accompaniment, making accessibility dependent partly on another household member’s availability.Qiang Wu participated in 27 accompanied events and Fang Wu in 16, including escort, shopping, leisure, and return-home travel.
- Evaluation scope: All scheduled activities were routable and completed, but the simulation primarily reveals differences in travel burden and coordination rather than severe service deprivation.No activity failed or required substitution, so the result does not establish universal neighborhood livability.
- Resident-agent evaluation: Simulated livability ratings reflected dispersed facilities and resident profiles, with healthcare access and trip chaining remaining constrained.The nearest major hospital was more than two kilometers away, while trip chaining received a mean rating of 5.2 and daily convenience, coordination, and recreation received ratings of 6.0-6.2.
- Resident-agent evaluation: Open-ended suggestions differed by resident profile, emphasizing accessible healthcare and walking conditions, childcare-supportive services, or parcel lockers and leisure destinations.These suggestions provide qualitative context for the resident-agent ratings.
5. Discussion
The prototype connects spatial context, household activity constraints, and resident-specific interpretation through an auditable hybrid simulation. Its demonstration shows that nominal facility availability can coexist with unequal travel and coordination burdens, while key limitations concern data accuracy, calibration, model sensitivity, and scope.
- The prototype combines spatial-KG-grounded LLM schedule generation and revision with rule-supported feasibility review, network materialization, and evidence-grounded resident interviews.
- Unlike conventional rule- or choice-model-based agents, the framework treats accessibility as an activity-based and household-dependent process rather than a static residential property.
- All scheduled activities were completed in the presented scenario, but mobility-sensitive residents had no trips within 15 minutes despite reaching required destinations.
- Facility availability did not guarantee usable access because practical burden depended on travel time, mobility capacity, activity timing, and other household members’ availability.
- Evidence-grounded interviews linked subjective convenience and burden interpretations to recorded routes, travel times, destinations, outcomes, and companionship, without substituting for resident surveys.
- Future applications could compare neighborhoods, test facility or transport interventions, and identify groups whose nominal proximity does not translate into manageable daily life.
- The prototype remains limited by OSM and sampled-home uncertainty, lack of trajectory calibration, sensitivity to LLM settings, and incomplete coverage of broader livability dimensions.
6. Conclusion
The study presents a spatial-KG-grounded LLM-agent prototype that extends neighborhood livability assessment beyond static proximity. Its demonstration identifies unequal travel and coordination burdens across resident groups and supports comparison, intervention testing, and identification of poorly served residents.
- The framework combines household-level LLM schedule generation and revision, rule-based feasibility review, network materialization, and evidence-grounded resident interviews.
- It connects structured spatial context with personalized activity-mobility simulation and traceable resident-specific evaluation.
- The prototype extends assessment beyond static facility proximity by considering mobility capacities, schedules, activity chains, and household responsibilities.
- The demonstration reveals unequal travel or coordination burdens when nominally accessible opportunities are considered across resident groups.
- The prototype can support neighborhood comparison, intervention testing, and identification of residents whose daily needs are poorly served.