Abstract
Social and behavioral interventions are critical for addressing societal challenges such as homelessness, disease, and poverty. However, these interventions often face limited resources and data. This research introduces "algorithmic social intervention" as a unified framework to enhance the targeting and delivery of these programs using AI and optimization. By focusing on decision-making under uncertainty, this work develops robust algorithms for influence maximization and submodular optimization. Applications in public health, specifically HIV prevention for homeless youth and tuberculosis control, demonstrate that these algorithmic approaches significantly outperform traditional heuristics in real-world settings.
Introduction: A Unified Framework
Algorithmic social intervention bridges the gap between theoretical algorithm design and the practical complexities of field deployments. The research trajectory follows a structured path from initial information gathering to the evaluation of field impact, ensuring that technical innovations are grounded in real-world needs.
The Challenge: Uncertainty and Resource Constraints
Governments and NGOs frequently deploy interventions under two major constraints: limited resources (budgetary and personnel) and limited data (incomplete knowledge of social networks or disease dynamics). Traditional AI models often assume a perfect "map" of the system, which is rarely available in the field. For example, in HIV prevention for homeless youth, social workers train "peer leaders" to spread awareness. However, mapping the entire social network of these youth is labor-intensive and often impossible.
Methodological Innovations
To address these challenges, several novel algorithmic families were developed:
Robust Influence Maximization
These algorithms handle uncertainty in how information spreads. DOSIM uses a game-theoretic approach, treating unknown spread probabilities as an adversary to find strategies that perform well even in worst-case scenarios. CHANGE is a "field-ready" evolution of this concept, designed to be computationally efficient and practical for use on standard laptops.
Exploratory Influence Maximization
The ARISEN algorithm identifies influential nodes without requiring a full graph. By exploiting the modular "community structure" of social networks and using random walks to estimate community sizes, it can find optimal "seeds" for information spread by querying only a tiny fraction of the population.
Robust Submodular Optimization
Many social problems exhibit diminishing returns, modeled as submodular functions. EQUATOR handles cases with multiple possible objective functions, scaling to large problems where previous methods failed. RASCAL focuses on risk-aversion, optimizing for the "Conditional Value at Risk" (CVaR) to prevent catastrophic failures in intervention outcomes.
Key Insights and Results
- Community Structure as a Proxy: You don't need a full map to find leaders. Estimating community sizes via random walks is sufficient to identify high-impact individuals.
- Robustness over Accuracy: Planning for the worst-case scenario (robust optimization) leads to more stable real-world outcomes than relying on "best guess" models.
- The "Laptop Test": Practical utility requires efficiency. The CHANGE algorithm runs in minutes on a standard laptop, making it accessible to social workers in the field.
Field Impact
- HIV Prevention: In pilot tests with homeless youth in Los Angeles, AI-based algorithms (CHANGE, DOSIM) reached 70-80% of the target population with health information, compared to only 27% reached by traditional status-quo methods (e.g., degree centrality).
- Data Efficiency: CHANGE achieved these results while surveying only 18% of the social network.
- Disease Control: Simulations for Tuberculosis in India suggested that the DOMO algorithm could avert over 8,000 cases per year compared to current policies.
- Computational Speed: EQUATOR solved robust optimization problems with 1,000 variables in under 40 seconds, whereas previous state-of-the-art methods timed out after 10 hours.
Future Directions
Future work aims to integrate multi-fidelity optimization, using simple surrogate models to guide complex simulations, and distributionally robust optimization to refine uncertainty models using available data samples. Additionally, accounting for the complexity of human behavioral responses to interventions remains a critical open challenge for the field.
Summary updated