SGP Digital Atlas

Features + Plexis Graph + Embeddings — Final Report

Propheus · April 2026 · 20 data sources · 5.98M population · 174,711 places

Features & Data

Atlas & Graph

Use Cases

Accuracy

Data Sources

Gov + commercial + satellite + live API

5.98M

Population

4.21M resident + 1.77M non-resident

637

Hex-8 Features

1,191 neighborhoods

612

Hex-9 Features

7,318 cells

114

Place Features

174,711 businesses

~449

Subzone Features

326 URA zones

Data Sources

Source	Records	Features derived
Overture Maps + OSM	174,711 places	79 place composition + 114 per-place
Overture Buildings	377,331	16 built environment
LTA Stations + Bus Stops	231 MRT + 44 LRT + 5,177 bus	18 transit + 8 GTFS + 19 place anchors
LTA Ridership (hourly)	12.3M taps/day	7 temporal transit (AM/PM/off/night)
Singapore GTFS 2026	230,914 trips	8 frequency features (headway by window)
OSM Road Network	550,991 segments	26 walkability (network walk from 214K-node graph)
SingStat Population	5,982,320	18 demographics + 12 dwelling type
HDB Resale Transactions	227,207	2 property price features
URA Master Plan	113,212 parcels	12 land use
NASA VIIRS Nightlights	2 epochs (2022+2024)	7 nightlight features (growth, commercial indicator)
WorldPop + WorldCover	Grid rasters	5 satellite (pop growth, land cover)
LTA DataMall API (live)	3K taxis + 2.6K carparks + 50K speed bands	10 dynamic (taxi, carpark, congestion, bus services)
LTA Bus Routes	26,711 route-stops (789 services)	Bus network topology + connectivity
Government amenity datasets	~3,000 (hawkers, clinics, parks, schools, hotels)	16 amenity counts + distance features
OSM POIs (4 layers)	52,317	4 supplementary counts

Feature Pillars — Hex-8 (637 features)

Pillar	#	Source	What it captures
Demographics	18	SingStat	Population total (5.98M), elderly %, non-resident, daytime intensity
Dwelling types	12	SingStat	HDB 1-5rm, condo, landed, by floor area
Built environment	16	Overture	Buildings, HDB blocks, floors, commercial/industrial
Land use	12	URA	Zoning %, entropy, fragmentation, dominant use
Transit	18	LTA	MRT/bus counts, taps (total + AM/PM/off/night)
GTFS frequency	8	GTFS 2026	Headway by time window, routes, departures
Walkability	26	OSM graph	Euclidean + network walk to 6 amenities, detour ratios
Amenities	16	Gov data	Hawkers, clinics, parks, supermarkets, hotels, schools
Place composition	79	Overture+OSM	24 categories, tiers, entropy, HHI, brands
Demand pull	12	Computed	6 pull scores (office/residential/transit/hotel/school/hawker)
Synergy	20	Computed	10 co-location scores (cafe×office, grocery×residential...)
Saturation	13	Computed	5 categories: supply/demand ratio + gap (pop_total denominator)
Satellite	12	VIIRS+WP+WC	Nightlight change, pop growth, land cover
Archetypes	15	K-means	6 types + 8 indices (vitality, accessibility, demand...) + 4 proxies
Micrograph	156	Pipeline	12 categories × 13 context vectors
Spatial context	123	H3 rings	Ring-1/2 max + pop-weighted aggregates
LTA dynamic	10	Live API	Taxi, carpark, speed, congestion, car dependency
Structure	8	Cross-scale	Interface, gradient, demand flow, ecosystem, self-containment
Property	2	HDB resale	Median PSF, transaction count

Place Features (114 per place)

Group	#	Key features
Identity	14	Category (24 types), price tier, branded, h3 keys
Competition	5	competitors_200m/500m, substitution_risk, market_share
Complementary	5	Cross-category diversity 300m, score
Anchors	19	14 types (MRT, bus, hawker, clinic, park, hotel, school, library, sports...)
Demand + synergy	18	6 pulls + demand_context + 10 synergies (target-category-only)
Transit	8	Network walk MRT/bus, GTFS headway, transit_score
Catchment	5	pop, elderly, nonresident, daytime intensity
Context	16+	Building, neighborhood, archetype, indices, nightlight
Supply-demand	5	saturation, demand_match, survivability_index

~200K

Graph Nodes

Feature-rich (114-637d)

1.49M

Graph Edges

39 relation types

128d

Embeddings

R-GCN two-head

Plexis

πλεξις

"weaving"

Four spatial levels + knowledge graph

Flat features answer WHAT (this hex has 42K pop, 0.2x cafe saturation). The graph answers WHY (because MRT funnels 125K commuters, adjacent hexes leak demand here, void-deck economy captures captive residents) and WHAT IF (remove the MRT → which places shift most?).

Level	Units	Features	Resolution	Role
Hex-9	7,318	612	~174m	Fine-grain context, walkability, micrograph
Hex-8	1,191	637	~461m	Primary: demand, gaps, archetypes, ecosystem
Subzone	326	~449	URA	Policy alignment
Places	174,711	114	Point	Competition, synergy, survivability
Plexis	~200K	1.49M edges	39 types	Structural reasoning, scenarios

Graph edge families

Family	Edges	Key relations	What it captures
Commercial	941,834	COMPETES, SYNERGIZES, SUBSTITUTES, EXIT_FRONTAGE, VOID_DECK	How businesses interact
Hierarchy	364,058	LOCATED_IN, IS_A, PARENT_OF, PART_OF	Containment + classification
Anchor	121,053	ANCHORED_BY, WALK_CATCHMENT, SERVES	Demand generators
Spatial	27,551	ADJACENT_TO, N/S/E/W_OF, ROAD, COASTAL	Physical connectivity
Structure	10,758	SAME_CLUSTER, LU_TRANSITION, DEVELOPMENT_FRONT	Urban evolution
Gradient	9,263	COMMERCIAL/HEIGHT/DENSITY/PRICE gradients	Change across space
Transit	5,770	CONNECTS_TO, FEEDS_INTO, SAME_CORRIDOR, EXPRESSWAY	Movement network
Supply-demand	5,260	UNDERSUPPLIED, OVERSUPPLIED, DEMAND_LEAKS, COMPARABLE	Gaps & opportunities

Two-head R-GCN embedding

Spatial Head (64d)

Encodes WHERE — trained on hierarchy + spatial + transit edges. Captures walkability, population, ecosystem.

Commercial Head (64d)

Encodes WHAT — trained on commercial + supply-demand edges. Captures category, competition, demand match.

Full (128d)

Concatenation of both. Used for similarity, expansion, anomaly detection, scenarios.

10 use cases powered by features + graph + embeddings

Each follows: start at entity → traverse typed edges → combine with flat features → produce ranked results + explanation.

1. Brand expansion

Build brand centroid from existing outlet embeddings → find hex-8 cells with similar graph structure but brand absent → rank by similarity × (1 − saturation).
Each brand gets different locations. Starbucks ≠ KFC ≠ Guardian.

2. Explain success

Traverse edges from any place: ANCHORED_BY MRT (transit demand), SYNERGIZES_WITH offices (lunch trade), VOID_DECK_OF HDB (captive demand). Graph tells the story.

3. Scenario simulation

"What if this MRT closes?" → Remove edges → recompute embeddings → places that shift most = most dependent. Quantifies infrastructure impact.

4. True food deserts

Follow UNDERSUPPLIED edges → check DEMAND_LEAKS_TO → no leak = true desert. Filters out CBD false positives.

5. Location-category fit

"Luxury restaurant here?" → Check COMPARABLE_TO (are comparables luxury areas?), PRICE_GRADIENT (gentrifying?), ANCHORED_BY hotels. Multi-signal reasoning.

6. Demand decomposition

Count incoming edges by type: ANCHORED_BY MRT = 35%, VOID_DECK = 25%, SYNERGIZES offices = 20%. Attribution, not just a number.

7. Anomaly detection

Place embedding vs hex embedding similarity. Low = structural misfit (retail in residential, bar in industrial). Flags risky locations.

8. Evolution tracking

DEVELOPMENT_FRONT + GRADIENT increasing + SAME_CLUSTER joining → "Tengah is the next Punggol." Growth prediction.

9. Competitive landscape

Traverse COMPETES_WITH → classify by synergy overlap: 46 threats + 12 allies + 27 substitutes. Net competitive position.

10. Cross-city transfer

Same relation schema for SGP + HKG → find HKG neighborhood matching Toa Payoh's graph structure. Universal urban patterns.

Live demo result: Bubble tea expansion

Top 3: Matilda/Punggol (36K pop, 125K transit, 0 BBT, sat 0.3x), Rivervale/Sengkang (43K pop, 0 BBT), Jurong West Central (48K pop, sat 0.2x). All Dense HDB, young families, massive undersupply. 1.6 seconds.

What the 128d Plexis embedding captures

R² = how much of each feature can be predicted from the 128d embedding alone. Higher = the graph encodes this information well.

Hex-8 R² (embedding → hex feature)

Feature	R²	Quality
walkability_score	0.898	EXCELLENT
pull_residential	0.894	EXCELLENT
ecosystem_completeness	0.827	STRONG
population	0.777	GOOD
pull_office	0.667	MODERATE
pc_total	0.661	MODERATE
transit_daily_taps	0.648	MODERATE

Place R² (embedding → place feature)

Feature	R²	Quality
anchor_score	0.908	EXCELLENT
demand_context_score	0.884	EXCELLENT
competitors_200m	0.774	GOOD
complementary_diversity	0.696	GOOD
transit_score	0.661	MODERATE
survivability_index	0.477	MODERATE

Other metrics

Metric	Value	What it means
Category classification	69.8%	Embedding alone predicts business type 70% of the time (24 categories)
Category separability	310x	Same-category places are 310x more similar than different-category in commercial head
Retrieval P@5	0.100	10% of top-5 nearest neighbors are same category (vs 4% baseline)
Archetype NMI	0.362	Embedding clusters partially recover pre-computed neighborhood archetypes
Link prediction Hits@10	14.1%	True connected node appears in top-10 predictions 14% of the time

Dynamic data adds +10% R²

Target	Static features only	+ LTA dynamic	Gain
ecosystem_completeness	0.715	0.818	+10.3%
saturation_fnb	0.479	0.524	+4.4%
idx_vitality	0.935	0.953	+1.8%

Summary by dimension

What embedding captures	R² range	Verdict
Structural / spatial (walkability, ecosystem, population)	0.78 – 0.90	EXCELLENT — these ARE graph properties
Demand (pull_residential, demand_context, anchor_score)	0.88 – 0.91	EXCELLENT — demand flows through edges
Competition (competitors, diversity)	0.70 – 0.77	GOOD — competition is graph density
Transit (taps, transit_score)	0.65 – 0.66	MODERATE — hub-dependent
Viability (survivability)	0.48	MODERATE — supply-side beyond graph
Category identity	69.8% acc, 310x sep	STRONG — commercial head works

Production formula: score = embedding_similarity × (1 − saturation) × demand_match
Features handle supply. Graph handles structure. Together they answer everything.

SGP Digital Atlas + Plexis · Propheus · April 2026 · 637+612+114 features · 1.49M graph edges · 128d R-GCN