Conscious infrastructure for fairness aware detection of household water insecurity in Hyderabad informal settlements

Technical Summary: Conscious Infrastructure for Fairness-Aware Detection of Household Water Insecurity in Hyderabad Informal Settlements

Problem Statement
Current applications of Artificial Intelligence (AI) and machine learning in urban water governance are predominantly organized around the operational perspective of utilities, focusing on network assets, flows, pressures, and infrastructure visibility. This orientation is analytically incomplete for informal settlements, where water insecurity is often experienced not as a total absence of infrastructure, but as the household burden of making unreliable or unsafe services usable. The paper identifies an "infrastructural visibility gap": the disjuncture between the burdens households experience (time loss, income pressure, health risks, coping costs) and the evidence that public decision systems recognize as actionable. When AI models are trained solely on utility-visible variables (e.g., source type, supply schedules), they risk optimizing for institutional legibility while missing households whose insecurity is expressed through labor and economic disruption. This constitutes a "problem-formulation bias" rather than merely a technical algorithmic bias.

Methodology
The study operationalizes a decision-support architecture termed Conscious Infrastructure Neural Systems (CINS). It is important to note that "neural" in this context refers to a linked decision architecture connecting household data, predictive classification, subgroup fairness audits, and institutional response protocols, not to the use of neural network algorithms.

• Data Source: The analysis utilizes a raw dataset comprising 1,700 household survey records across 34 informal settlements in Hyderabad, India, and 34 water-quality sample records. The data originates from doctoral research on drinking-water access and decision-making.
• Outcome Variable: The primary outcome is a binary indicator of whether a household reported that its work schedule was disrupted by drinking-water supply or availability timing. This serves as a proxy for "binding" water insecurity, capturing direct economic and temporal consequences rather than subjective perception.
• Model Comparison: Two logistic regression models were estimated and compared on a held-out test set (30% of data, stratified):
  1. Utility-Visible Model: Includes variables closer to conventional infrastructure knowledge (source use, supply timing, storage duration, settlement ID, linked microbiological risk).
  2. Household-Burden Model: Includes all utility-visible variables plus household-burden variables (income/expenditure midpoints, expenditure ratios, perceived quality/quantity ratings, illness/treatment indicators, fetch burden, and multi-source dependence).
• Evaluation Metrics: Performance was assessed using accuracy, precision, recall, F1-score, AUC, and specifically the False Negative Rate (FNR). The study prioritizes the reduction of FNR among vulnerable subgroups (lowest-income quartile, multi-source households, tanker-dependent households, and those in microbiological-risk settlements) over aggregate accuracy.
• Fairness Audit: Subgroup analysis was conducted to determine if the inclusion of burden variables reduced under-detection for specific vulnerable groups. A Random Forest feature-importance analysis was used as a supplementary interpretability check.

Key Results

• Model Performance: The household-burden model outperformed the utility-visible model on the held-out test set.
  • Accuracy: Increased from 0.669 to 0.731.
  • Recall: Increased from 0.636 to 0.714.
  • F1-Score: Increased from 0.513 to 0.593.
  • AUC: Increased from 0.731 to 0.818.
  • False Negative Rate: Decreased from 0.364 to 0.286.
• Confusion Matrix: The household-burden model correctly identified 100 true positives (disrupted households) compared to 89 in the utility-visible model, reducing missed disrupted households from 51 to 40.
• Subgroup Fairness:
  • Improvements: The FNR decreased significantly for the lowest-income quartile (32.1% to 25.0%) and multi-source households (40.5% to 31.6%).
  • Anomaly: The FNR for tanker-dependent households increased slightly (8.8% to 12.5%). The authors interpret this as a result of tanker use already being a strong predictor in the utility-visible model; adding burden variables redistributed model weights, slightly weakening performance for this specific subgroup whose risk was already highly visible.
• Feature Importance: In the household-burden model, economic pressure (expenditure, income margin), perceived service quality, and fetch burden emerged as high-ranking predictors alongside infrastructure variables, confirming that household experience adds decision-relevant information.

Key Contributions

1. Conceptual Framework: The paper introduces the infrastructural visibility gap as a bridge between household water insecurity, data justice, and urban governance. It defines this gap as the difference between household-experienced burdens and the evidence recognized by decision systems.
2. Empirical Evidence: It provides a reproducible comparison demonstrating that incorporating household-burden variables improves the detection of high-risk households and reduces false negatives among vulnerable subgroups, challenging the sufficiency of infrastructure-centric data.
3. Governance Architecture (CINS): The paper proposes CINS not merely as a predictive tool but as a governance architecture that links predictive classification to institutional response protocols. It argues that fairness-aware AI must be designed around the detection of household burdens and must trigger specific actions (e.g., retesting, supply correction, affordability assessment) rather than remaining as passive dashboard outputs.
4. Reframing Fairness: The study shifts the evaluation of fairness in infrastructure AI from aggregate accuracy to the distribution of false negatives across vulnerable subgroups, arguing that fairness is a property of data inclusion and problem formulation, not just post-hoc model adjustment.

Significance and Claims
The paper claims that fairness in infrastructure AI is not achieved merely by correcting model outputs but by fundamentally altering the data architecture to include household burdens. The central argument is that a household can be "formally visible" (connected to a network) yet "analytically invisible" regarding the labor and economic burdens it bears.

The authors modestly claim that their analysis demonstrates that household-burden variables improve detection under a transparent baseline specification. They do not claim to have identified a final optimal model or to have fully deployed CINS operationally. Instead, the paper positions the model comparison as a diagnostic instrument for institutional visibility, showing what is missed when household burden is excluded. The significance lies in the assertion that for AI to be fairness-aware in urban water governance, it must be designed to recognize the lived realities of informal settlements—specifically service fragmentation and the associated burdens—rather than relying solely on infrastructure visibility. The paper concludes that without integrating household-burden data and linking predictions to response protocols, AI risks making water governance more efficient while preserving the invisibility of those most affected by unreliable access.