Meta-Retrieval: Personalizing Prototype Retrieval Strategies Using Contextual Bandit Learning

Sekar & Wang (2018) introduce contextual bandits for model selection in similarity retrieval, but this concept hasn’t been widely adopted in CBDS. Most systems use a fixed similarity metric or retrieval algorithm, ignoring user diversity and situational nuance. This research proposes a meta-retrieval layer: as users interact with the CBDS, a contextual bandit agent observes their context (user expertise, urgency, decision type, feedback) and dynamically chooses among multiple retrieval models—e.g., dense vs. sparse, knowledge-based vs. data-driven, or causal vs. correlational. Over time, the system learns which strategies perform best for which contexts, continually optimizing for user satisfaction, accuracy, or speed. This approach could yield more adaptive, user-centered CBDS systems, especially valuable in high-variance domains like healthcare, maintenance, or urban planning (as seen in Koornneef et al., 2021; Shulajkovska et al., 2024). It’s a step towards “intelligent retrieval selection,” rather than static retrieval, and could significantly boost system effectiveness.

References:

1. A Web-Based Decision Support System for Aircraft Dispatch and Maintenance. H. Koornneef, W. Verhagen, R. Curran (2021). Aerospace.
2. Case-Based Decision Support System with Contextual Bandits Learning for Similarity Retrieval Model Selection. B. Sekar, Hui Wang (2018). Knowledge Science, Engineering and Management.
3. Artificial Intelligence-Based Decision Support System for Sustainable Urban Mobility. Miljana Shulajkovska, Maj Smerkol, Gjorgji Noveski, Marko Bohanec, M. Gams (2024). Electronics.