Co‑Adaptive Eco‑Nudging: A Privacy‑Preserving Contextual Bandit with User‑Taught Preferences in Everyday Browsing

Paper Title

Co-Adaptive Eco-Nudging: A Privacy-Preserving Contextual Bandit with User-Taught Preferences in Everyday Browsing

Publication Info

• Topic area: Sustainable Human-Computer Interaction (SHCI) and digital eco-feedback systems.
• Keywords: Eco-nudges, contextual bandits, privacy-preserving AI, user autonomy, energy ROI, ethical efficacy frontier, personalization, digital sustainability, human-centered ML, net environmental impact.

Background and Problem

• Problem / challenge: Digital eco-nudges often lack rigorous evaluation frameworks, fail to balance efficacy with user autonomy, and rarely account for net environmental impacts including system overheads. Personalization methods are frequently opaque and server-based, raising privacy concerns.
• Significance: Addressing these gaps is critical to advancing sustainable HCI practices that are both effective and respectful, ensuring measurable environmental benefits while maintaining user trust and autonomy.
• Motivation and related work: Prior research highlights the potential of eco-feedback and digital nudging but suffers from evaluation confounds (e.g., mismatched message content/length or delivery budgets) and short study horizons. Contextual bandits offer a promising method for adaptive personalization but are rarely applied in privacy-preserving, on-device settings. Ethical concerns about autonomy, transparency, and privacy remain underexplored in the context of digital sustainability interventions.

Solution

• Proposed approach: A privacy-preserving, co-adaptive contextual bandit system integrated into a browser extension, designed to learn user preferences and optimize eco-nudging interventions while respecting autonomy constraints.
• Novelty:
  1. Parity-controlled evaluation of eco-nudges using opportunity-based denominators to reduce confounds.
  2. Introduction of a constrained contextual bandit that incorporates DoNothing actions, user-governed quiet hours, and explicit feedback.
  3. Development of the Ethical–Efficacy Frontier (EEF) to visualize autonomy–compliance trade-offs and an Energy ROI framework to estimate net environmental benefits.
  4. Empirical evidence on routine digital behaviors (tabs, streaming, printing, large transfers) with implications for sustainable HCI.
• Procedure and key techniques:
  • Study 1 compared context-tailored prompts to generic and control conditions under strict parity of message content and delivery budgets.
  • Study 2 evaluated a co-adaptive, on-device contextual bandit against a rule-based policy, incorporating user feedback, autonomy constraints, and energy-efficient design principles.
  • Both studies employed rigorous instrumentation, mixed-effects models, and sensitivity analyses to ensure validity and reproducibility.

Results

• Concrete findings:
  • Context-tailored prompts improved compliance for streaming (+8.0 pp) and tabs (+7.7 pp) compared to generic prompts (Study 1).
  • The co-adaptive bandit achieved higher compliance for streaming (+6.9 pp) and tabs (+7.2 pp) compared to the rule-based policy, while maintaining slightly higher autonomy (Study 2).
  • Energy ROI analysis indicated net positive environmental benefits under reasonable assumptions, with streaming and tabs contributing the largest savings.
• Advantage over baselines:
  • Context-tailored prompts outperformed generic prompts and control conditions.
  • The co-adaptive bandit demonstrated learned restraint (frequent selection of DoNothing) and better temporal alignment, shifting the Ethical–Efficacy Frontier outward.
• Experiments / evaluation:
  • Study 1: N = 178 participants, randomized into three arms (Control, Generic, Context-Tailored) over two weeks.
  • Study 2: N = 54 participants, comparing Bandit vs Rule-based policies over a two-week intervention, two-week withdrawal, and one-week follow-up.
  • Metrics included opportunity-level compliance, autonomy, trust, privacy concerns, and energy ROI.
• Limitations and future work:
  • Short intervention duration limits claims about long-term habit formation and rebound effects.
  • Energy ROI relies on literature-calibrated proxies rather than direct power metering.
  • Results may not generalize to mobile platforms or different cultural contexts. Future work should explore longer deployments, richer affordances for high-effort behaviors, and randomized prompt intensity.

Summary

This paper introduces a privacy-preserving, co-adaptive contextual bandit system for eco-nudging, demonstrating improved compliance and autonomy over rule-based policies. Two field studies show that minimal factual tailoring and adaptive timing significantly enhance user engagement in routine digital behaviors such as streaming and tab management. The Ethical–Efficacy Frontier and Energy ROI frameworks provide new lenses for evaluating the trade-offs between efficacy, autonomy, and environmental impact. While short-term results are promising, longer deployments and direct energy instrumentation are needed to assess long-term sustainability and scalability. These findings contribute to sustainable HCI by aligning personalization, privacy, and net-positive environmental outcomes.