In many real-world machine learning applications, particularly in resource-constrained domains such as medical diagnostics, acquiring feature values is a costly and often sequential process. Active Feature Acquisition (AFA) addresses this by selecting the most informative subset of features to acquire for a given instance, balancing predictive accuracy against acquisition costs. Conventional AFA strategies often rely on static, global feature importance metrics, which are instance-agnostic and can be suboptimal when feature relevance is context-dependent. We investigate a practically relevant two-stage acquisition scenario, where an initial subset of features is observed, and a subsequent non-overlapping subset is selected for acquisition. We propose a novel AFA strategy that leverages instance-specific model explanations, specifically Shapley additive explanations (SHAP), to guide the selection process. We conduct a systematic comparative study, evaluating three distinct acquisition strategies: random acquisition (as a baseline), acquisition guided by static global feature importance lists, and the proposed SHAP-based approach. Using XGBoost-trained models, these methods are evaluated across ten benchmark datasets under two missingness scenarios: Missing Completely at Random (MCAR) and Missing Not at Random (MNAR). The empirical results demonstrate that the SHAP-based strategy significantly outperforms static global feature importance methods in complex settings, particularly under MNAR missingness, where the context of observed features is critical for effective decision-making. While it also performs strongly in simpler MCAR scenarios (especially at high missingness rates), its robustness in more realistic settings suggests that utilizing instance-specific explanations provides a powerful and adaptive mechanism for personalized and effective feature acquisition.
|
