A Survey on Active Feature Acquisition Strategies

Active feature acquisition (AFA) studies how to sequentially acquire features for each data instance to trade off predictive performance against acquisition cost. This survey offers the first unified treatment of AFA via an explicit partially observable Markov decision process (POMDP) formulation. We place this formulation in the broader literature on optimal information acquisition and, more specifically, in a family of structured POMDPs (for example, information-gathering and sensing POMDPs) whose assumptions and algorithmic tools directly apply to AFA. This connection provides a common language for comparing problem settings and methods, and it highlights where AFA can leverage established results in structured POMDP planning and approximation. Building on this perspective, we present an up-to-date taxonomy of AFA methods that (roughly) mirrors standard approaches to solving POMDPs: (i) embedded cost-aware predictors (notably cost-sensitive decision trees and ensembles), (ii) model-based methods that plan using learned probabilistic components, (iii) model-free methods that learn acquisition policies from simulated episodes, and (iv) hybrid methods that combine the strengths of model-based and model-free approaches. We argue that this POMDP-centric view clarifies connections among existing methods and motivates more principled algorithm design. Since much prior work is heuristic and lacks formal guarantees, we also outline routes to guarantees by connecting AFA to adaptive stochastic optimization. We conclude by highlighting open challenges and promising directions for future research.

💡 Research Summary

This survey presents a comprehensive treatment of Active Feature Acquisition (AFA), a problem where an agent must decide, for each data instance, which costly features to acquire sequentially and when to stop and make a prediction. The authors formalize AFA as a finite‑horizon, episodic Partially Observable Markov Decision Process (POMDP). In this formulation, the state consists of the set of already observed features, their values, the hidden remaining features, and the true label. Actions are either the acquisition of a yet‑unobserved feature (subject to a per‑instance budget) or a STOP action that triggers prediction. Transitions and observations are deterministic; the belief state is the posterior distribution over the hidden features and label given the observed ones. Immediate rewards penalize feature costs (‑α c_a) while the STOP action yields a terminal reward equal to the negative prediction loss (‑ℓ). Consequently, maximizing expected cumulative reward is exactly the trade‑off between prediction accuracy and acquisition cost expressed in the standard AFA objective.

Using this POMDP lens, the authors reorganize the fragmented AFA literature into four broad categories that mirror classic POMDP solution strategies:

1. Embedded cost‑aware predictors – decision trees, ensembles, or other models that internalize acquisition costs; a trained tree can be viewed as a deterministic policy that stops at leaf nodes.
2. Model‑based methods – learn probabilistic components (e.g., p(x_U | x_S), p(y | x_S)) and then apply model‑based planning algorithms such as value iteration, point‑based solvers, or Monte‑Carlo Tree Search.
3. Model‑free methods – treat the acquisition process as a reinforcement‑learning problem, learning a policy or Q‑function directly from simulated episodes using deep Q‑networks, policy gradients, actor‑critic, etc.
4. Hybrid methods – combine the strengths of the above, for example by initializing a policy with a model‑based planner and refining it with model‑free RL, or by augmenting a cost‑aware tree with learned look‑ahead.

The survey connects AFA to a broader family of optimal information‑gathering problems, including adaptive stochastic optimization, submodular value‑of‑information heuristics, sequential experimental design, active learning, multi‑armed bandits, and robotics perception. It highlights that many classic systems—such as medical expert systems (e.g., Pathfinder) and cost‑sensitive decision trees—are early instances of AFA, and that modern deep predictors can be seamlessly integrated into the POMDP framework via masking, imputation, or deep set encoders.

A significant contribution is the discussion of theoretical guarantees. By recognizing AFA as an adaptive stochastic optimization problem, the authors point out that myopic value‑of‑information policies enjoy (1 − 1/e) approximation under submodularity, and that structured POMDP planners can inherit similar bounds. However, most existing AFA work remains heuristic, lacking formal performance guarantees, especially when dealing with approximate belief updates or learned models.

The paper also surveys practical considerations: training typically assumes fully observed data, with the predictor either pretrained on random masks or jointly learned with the acquisition policy. Online, cost‑aware training—where features are acquired during learning—is still rare. Computational challenges arise from belief‑state updates in high‑dimensional spaces and from online planning under strict latency constraints.

Finally, the authors outline open challenges and future directions: scalable belief‑state approximation, uncertainty‑aware modeling (e.g., Bayesian neural networks), online cost‑sensitive learning, stronger approximation guarantees beyond myopic heuristics, transfer and multi‑task acquisition policies, and integration with privacy‑preserving mechanisms.

In summary, by casting AFA as a POMDP, the survey unifies disparate methods under a common mathematical language, clarifies their relationships to classic decision‑theoretic frameworks, and provides a roadmap for developing principled, scalable, and theoretically grounded active feature acquisition systems.