Title: Learning-To-Measure: In-Context Active Feature Acquisition

Abstract: Active feature acquisition (AFA) is a sequential decision-making challenge aimed at enhancing model performance on test instances through the adaptive selection of features to acquire. However, practical applications of AFA frequently rely on retrospective data characterized by systematic feature missingness and a scarcity of task-specific labels. Existing approaches typically focus on acquisition for a single, fixed task, which restricts their scalability. To overcome this constraint, we define the meta-AFA problem, which seeks to learn acquisition policies applicable across a diverse range of tasks. We propose Learning-to-Measure (L2M), a framework built upon two core components: robust uncertainty quantification for unseen tasks and an uncertainty-driven greedy feature acquisition agent designed to maximize conditional mutual information. L2M leverages a sequence-modeling or autoregressive pre-training strategy to ensure reliable uncertainty estimates, even for tasks with arbitrary patterns of missing data. By operating directly on datasets with retrospective missingness, L2M executes the meta-AFA task in-context, thereby removing the need for per-task retraining. Evaluations on both synthetic and real-world tabular benchmarks reveal that L2M performs on par with or better than task-specific baselines, especially in scenarios involving high levels of missingness and limited labels.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC