Graph-based supply-chain optimisation
From network structure to practical scenario evidence
This prototype connects the thesis network analysis, predictive models and Monte Carlo revenue simulation in one transparent decision-support workflow.
Synthetic study data · indicative results · not operational advice
01
Map
Identify access constraints
02
Model
Measure added information
03
Explore
Test bounded assumptions
Executed analysis
Reproducible snapshot
Headline evidence
Decision pathway
How to use the evidence
-
1
Find a question
Use route exposure and centrality to identify where closer operational investigation may be useful.
-
2
Check the evidence
Compare model specifications and feature importance before interpreting structural signals.
-
3
Test assumptions
Use the Scenario Lab to inspect conditional implications, not guaranteed outcomes.
Interpretation boundary
What this prototype is — and is not
Technical proof of concept
Integrates the executed thesis outputs into an accessible interface.
Transparent scenario tool
Shows the assumptions and sensitivity grid behind each estimate.
Not a field estimate
All outputs are based on a synthetic analytical sample.
Not an automated decision
Human judgement and observed-data validation remain essential.
Network intelligence
Illustrative topology
Explore supply-chain structure
Inspect structural indicators, route exposure and critical intermediaries in the executed synthetic network.
Deterministic study view
Directed market network
Farmers
Buying points
Depots
Ports
Upstream farmersMarket flow →Ports
The topology is a deterministic visual summary, not a geographic map. Exact operational use requires validated coordinates and road-network routing.
Weighted betweenness
Critical intermediary ranking
| Rank | Node | Type | Region | Betweenness | Degree | Select |
|---|
No intermediary matches the current filters.
Reachable farmers: 900
Route-cost distribution
128.10Minimum
308.10Median
452.2590th %ile
648.80Maximum
Question for field investigation
Which farmers consistently face route costs above 452.25?
Observed transport, road quality and sale-location data would be needed to answer this responsibly.
Predictive analytics
5-fold cross-validation · n = 900
Compare model evidence
Review cross-validated performance with and without graph-derived features across three evaluation metrics.
Model comparison
Cross-validated R²
Base featuresBase + network
Augmented XGBoost model
Feature importance
Evidence boundary
Read the gain correctly
No target leakage in graph construction
Route cost and network features were derived from graph structure without using the price target.
Injected signal, not causal proof
The synthetic price generator intentionally contains route-cost and community effects. The gain demonstrates signal recovery.
Yield R² · base0.4870
Yield R² · + network0.4769
The yield check does not show a comparable improvement, supporting the intended specificity of the synthetic price signal.
Full diagnostics
Cross-validated model results
| Model | Features | RMSE | MAE | R² |
|---|
Monte Carlo revenue analysis
Scenario, not forecast
Test collaborative transport assumptions
Explore the executed 3 × 3 sensitivity grid and a clearly labelled participation rescaling.
Estimated aggregate revenue gain
8.87%
Central executed scenario
Mean baseline revenueGHS 9,153.69per farmer
Mean optimised revenueGHS 9,965.87per farmer
95% simulation interval8.59%–9.14%central simulation
Executed sensitivity
Click a cell to apply
Mean revenue gain
Central simulation
Gain distribution
8.59%Lower 95%8.87%Mean9.14%Upper 95%
Working comparison
Saved scenarios
| Scenario | Saving | Premium | Participation | Gain | Mean revenue | Remove |
|---|
No saved scenarios. Add the current result to compare alternatives.
Scenario, not forecastAssumptions do not guarantee returns.
Participation is heterogeneousThe executed model caps probability at 90%.
Field validation requiredObserved cost, price and adoption data are essential.
Reproducibility & governance
Checking service…
Audit the method and its limits
Trace every headline result to the executed settings and review the safeguards required before field use.
Graph construction
Route-weight settings
- Random seed
- 20250115
- Distance coefficient α
- 1.0
- Tariff coefficient β
- 0.8
- Time coefficient γ
- 0.6
- Reliability coefficient δ
- 25.0
- Louvain resolution
- 1.0
- Route algorithm
- Dijkstra
Model evaluation
Predictive settings
- Cross-validation
- k = 5, shuffled
- Random Forest
- 400 trees
- RF maximum depth
- None
- XGBoost
- 500 trees
- XGB depth
- 4
- Learning rate
- 0.05
- Target
- GHS/kg
Revenue simulation
Monte Carlo settings
- Iterations
- N = 5,000
- Central saving
- 35%
- Central premium
- GHS 0.30/kg
- Mean gain
- 8.87%
- Standard deviation
- 0.14
- 95% interval
- 8.59%–9.14%
- Sensitivity span
- 4.88%–12.86%
Data flow
What the service does
1Frozen JSONVerified thesis output
→
2Read-only serviceNo uploads or database
→
3Browser viewsBounded exploration
GET /Interface200GET /api/healthOperating mode…GET /api/resultsFrozen results…GET /api/scenarioBounded lookup…Deployment boundary
Required before operational use
- 01Observed-data partnership
Agree a data dictionary, governance protocol and institutional responsibilities.
- 02External and temporal validation
Re-estimate the graph and models using observed multi-season data.
- 03Human-centred research
Conduct accessibility, terminology and task-based studies with intended users.
- 04Security and accountability
Add role-based access, encryption, audit logging, monitoring and incident response.
Interpretation guide
Frequently asked questions
Can the prototype identify a depot that should be removed?
No. Betweenness identifies candidates for review, but the thesis did not run removal-induced fragmentation experiments. Intervention requires observed network evidence and resilience testing.
Does the 8.87% result predict what farmers will earn?
No. It is the mean result of a synthetic, assumption-dependent Monte Carlo scenario. It is not a forecast, guarantee or observed sector estimate.
Do network features prove a causal effect on price?
No. Their predictive contribution recovers a structural signal deliberately injected into the synthetic price-generating process. Field causal evidence is still required.
Has the interface been tested with farmers or cooperative leaders?
No human-participant usability study was conducted. The implemented tasks were verified by the developer as functional acceptance tests only.
Does the prototype store personal information?
No. It uses a frozen synthetic snapshot, accepts no uploads and stores no farmer or user data. Saved scenario comparisons exist only in the current browser session.
Need the underlying evidence?Download the exact JSON snapshot or print a concise prototype record.