Datasets:

greenadntan
/

wom-platform-romania-moldova

What it does: Maps where brand advocacy clusters geographically, identifies natural social communities, simulates how word-of-mouth spreads through those communities, and scores every ~5 km² hex cell for marketing amplification potential.

Architecture

┌─────────────────────────────────────────────────────────┐
│                    CLIENT INTERFACE                       │
│  Gradio Dashboard + API                                  │
│  • Upload CRM data → View propensity heatmaps            │
│  • Get activation recommendations per community          │
└──────────────┬──────────────────────────────┬────────────┘
               │                              │
    ┌──────────▼──────────┐       ┌───────────▼───────────┐
    │  SCORING ENGINE     │       │  ACTIVATION ENGINE     │
    │  Per-cell propensity│       │  Budget optimization   │
    │  scoring (XGBoost + │       │  Channel recommendation│
    │  graph features)    │       │  (Influence Max)       │
    └──────────┬──────────┘       └───────────┬───────────┘
               │                              │
    ┌──────────▼──────────────────────────────▼───────────┐
    │              DIFFUSION MODEL LAYER                   │
    │  • SIR-based brand adoption propagation model        │
    │  • Geographic community graph                        │
    │  • Temporal spread prediction                        │
    └──────────┬──────────────────────────────┬───────────┘
               │                              │
    ┌──────────▼──────────┐       ┌───────────▼───────────┐
    │  GEO-FEATURE LAYER  │       │  COMMUNITY GRAPH       │
    │  Per H3-cell:       │       │  Nodes = H3 cells     │
    │  • OSM features     │       │  Edges = similarity   │
    │  • Census demographics│     │  • Commuting flows    │
    │  • Satellite-derived│       │  • School catchments  │
    │  • POI density      │       │  • Church parishes    │
    │  • GeoLLM embeddings│       │  • Shopping patterns  │
    └─────────────────────┘       └───────────────────────┘

How It Works

1. H3 Hexagonal Grid

58,670 cells at H3 resolution 7 (~5.44 km² each)
Uniform coverage of Romania (55,856 cells) + Moldova (2,814 cells)
Each cell has 6 neighbors → natural graph structure

2. Feature Engineering (per cell)

OSM-derived: shops, restaurants, schools, churches, bus stops, hospitals, banks, etc.
Demographic estimates: population, urbanization score, income index, education, internet penetration
Community metrics: social cohesion, diaspora connection strength
Spatial: distance to nearest city, number of cities within 50km

3. Community Detection

Geographic proximity graph weighted by socioeconomic similarity
Louvain algorithm identifies natural "word-of-mouth communities"
Graph centrality metrics (degree, betweenness, clustering coefficient)

4. SIR Diffusion Model

Susceptible → Infected → Recovered epidemiological model
Brand WOM spreads through community graph
Infection rate modulated by: community cohesion × internet penetration
Identifies "susceptible zones" ready for brand entry

5. Propensity Scoring (0-100)

Composite score from:

Community cohesion (25%) — high cohesion amplifies WOM
Population density (20%) — larger audience
Internet penetration (15%) — digital WOM speed
Graph betweenness centrality (15%) — well-connected nodes
Clustering coefficient (10%) — tightly-knit communities
Susceptibility state (15%) — not yet saturated

Data Sources

Layer	Romania 🇷🇴	Moldova 🇲🇩
Census	INS 2021 (comuna/sat level)	BNS 2024 (fresh!)
Map Data	OpenStreetMap (excellent coverage)	OpenStreetMap (good)
Boundaries	ANCPI / Eurostat GISCO	geo.gov.md / Eurostat
Satellite	Sentinel-2 (ESA, free)	Sentinel-2
Business Registry	ONRC (~2M companies)	ASP
Social Media	Facebook (~11M), Instagram, TikTok	Facebook (~1.5M), TikTok

Key Romania/Moldova Advantages

Opportunity	Why It Matters
Diaspora Effect	~5M Romanians abroad → remittance corridors create unique WOM pathways
2024 Moldova Census	Freshest census data in the region — demographic precision advantage
Fast Internet	Romania has among the fastest broadband in Europe → digital WOM is highly trackable
Church Networks	Strong Orthodox parish structure → natural community boundaries
EU Data Access	Eurostat, Copernicus, INSPIRE directive → rich open geospatial data
First Mover	No Herdify competitor exists in RO/MD

Files

File	Description
`build_mvp.py`	Complete pipeline: grid → features → graph → diffusion → scoring → map
`app.py`	Gradio dashboard for interactive exploration
`README.md`	This file

Quick Start

pip install h3 numpy pandas networkx python-louvain scipy folium pyarrow gradio

# Run the full pipeline
python build_mvp.py

# Launch interactive dashboard
python app.py

Revenue Model (Indicative)

Tier	What's Included	Price
Explorer	Static propensity map, 1 brand, quarterly	€500-1,000/mo
Growth	Dynamic map, community insights, activation recs	€2,000-5,000/mo
Enterprise	API, real-time CRM, multi-brand, custom models	€5,000-15,000/mo

Target customers: eMAG, Kaufland, Lidl Romania, Orange, Vodafone, ING, BCR, UniCredit

Academic Foundations

GeoLLM (2310.06213): OSM + LLM knowledge achieves r²=0.72 for socioeconomic prediction
DeepIM (2305.02200): GNN-based influence maximization
NNN (2504.06212): Transformer-based marketing mix with geographic embeddings
SIR Models: Epidemiological diffusion adapted for brand propagation

Production Roadmap

H3 grid covering RO + MD
Feature engineering pipeline
Community detection (Louvain)
SIR diffusion simulation
Propensity scoring model
Interactive map generation
Real OSM Overpass extraction (replace modeled features)
Romania 2021 Census microdata integration
Moldova 2024 Census integration
Sentinel-2 satellite features (NDVI, nightlights, built-up)
Real client CRM data calibration
Multi-brand diffusion (different β/γ per brand)
Diaspora network layer
Gradio production dashboard with live scoring API
A/B testing framework for causal validation

Downloads last month: 31

Number of rows:

Total file size:

58.3 kB