Case Study: NGO Recommender System for Equitable Resource Distribution
– Context: humanitarian NGO serving marginalized communities
– Data constraints: sensitive, sovereign, possibly incomplete
– Model constraints: fairness-aware, energy-efficient, explainable
– Deployment decisions: cloud vs. edge; jurisdictional limitations
– Discussion on socio-technical trade-offs and ethical dilemmas
1. Context
You are part of a data science team working with a humanitarian NGO that supports marginalized and underserved communities in the Global South. The organization is building a recommender system to suggest educational, health, and economic support services to individuals based on demographic profiles, local context, and historical access patterns.
Your goal is to design and prototype a responsible and fair recommendation system that is technically sound while respecting data sovereignty, inclusivity, and environmental sustainability.
2. Objectives of the Hands-On
- Build or simulate a prototype recommender using Python (scikit-learn or similar).
- Apply fairness metrics to evaluate and mitigate potential bias.
- Measure energy impact using tracking tools like
codecarbon. - Make a deployment recommendation based on jurisdictional, environmental, and equity trade-offs.
- Reflect on the ethical dilemmas and power dynamics embedded in your design.
3. Considerations and constraints
3.1 Data Constraints
- Sensitive data: includes ethnicity, location, health status, socio-economic indicators.
- Sovereign data: stored on regional servers due to local or Indigenous data governance laws (e.g., OCAP, GDPR).
- Incomplete or biased: certain groups (e.g., Indigenous communities, women, undocumented migrants) are underrepresented due to historical inequalities or collection barriers.
3.2 Model Constraints
- Must be fairness-aware: avoid reinforcing existing inequities (e.g., services only targeting urban or male populations).
- Must be energy-efficient: the NGO operates in low-resource settings with limited access to cloud infrastructure.
- Must be explainable: users and NGO staff need to understand why recommendations are made (e.g., for trust and compliance).
3.3 Deployment Considerations
- Cloud options may offer compute resources but violate sovereignty or sustainability goals.
- Edge computing can reduce latency and emissions but limits model complexity.
- Hybrid architectures (e.g., local + cloud fallback) may offer compromises but introduce complexity.
- Jurisdictional constraints: hosting in EU or Africa versus U.S. (legal exposure to non-local law enforcement or surveillance).
4. To-Do Instructions
The general lines of the solution can be found here
- 1. Data Simulation
- 2. Build a Simple Recommender
- 3. Evaluate Fairness
- 4. Track Energy Consumption
- 5. Deployment Trade-Off Matrix
Discuss in your group: What is the most responsible deployment strategy? We will discuss my perspective in class.
| Criteria | Cloud (EU) | Sovereign Local | Edge Device |
|---|---|---|---|
| Compute Power | High | Medium | Low |
| Legal Compliance | Partial | High | Medium |
| Energy Footprint | High | Medium | Low |
| Latency | High | Low | Very Low |
| Inclusivity Control | Low | High | High |
| Deployment Complexity | Medium | High | Medium |
4. Wrap-Up Discussion Prompts
- What biases emerged from the data and the model? How did you address them?
- How did energy tracking influence your modeling decisions?
- Which deployment option balances fairness, sovereignty, and sustainability?
- What epistemic or power asymmetries are still embedded in your pipeline?
- How would you document your decision for an ethics audit?
Optional Deliverables (Extension)
- Reflective mini-essay or blog post: “How can I decolonize and decarbonize my pipeline?”
- A 1-page reflection: “How I balanced sovereignty, fairness, and sustainability in my design”
- Fairness-aware scoring prototype or resource optimization policy sketch
- A JSON or YAML model card including:
- Dataset description
- Fairness audit summary
- Carbon estimate
- Deployment recommendation