Traversal-Based Action Shortlisting
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Selecting high-value actions from very large candidate sets is a common problem in decision systems, recommendation engines, and optimization pipelines. The standard approach is global ranking, where all actions are scored using a fixed weighted function and the top candidates are selected. However, this approach degrades under distribution shift, feature noise, and situations where optimal actions depend on interactions rather than dominant features. This work introduces a traversal-based selection mechanism that replaces global scoring with local exploration on a structured graph. Instead of evaluating all candidates, the method moves through the action space using constrained one-step transitions and selects candidates from a small local neighborhood using robust statistics. A lightweight targeted probe monitors a single systematically under-exposed region to ensure recovery of sparse optimal actions. Across action spaces of up to 200,000 candidates, the method improves top-1 recall under weight shift by +27.5 percentage points and under feature noise by +9.9 percentage points compared to weighted ranking baselines. It also achieves perfect recovery of sparse optima with negligible additional cost. Shortlist sizes remain below 0.02% of the total action space, demonstrating strong compression with minimal loss of performance. These results show that local traversal combined with targeted exposure provides a robust alternative to global ranking, particularly in settings where the underlying utility function is unstable or partially unknown.
Fields of Study & Subjects
Recommender Systems and Techniques
Advanced Bandit Algorithms Research
Information Retrieval and Search Behavior
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 29 March 2026
- Language
- EN
- Document type
- Preprint