Dynamic Feature Selection (DFS)

• Dynamic Feature Selection (DFS) is an adaptive approach that sequentially acquires relevant features based on current observations and cost constraints.
• It employs methods like greedy conditional mutual information and reinforcement learning to balance predictive accuracy with reduced feature acquisition costs.
• Recent research in DFS integrates specialized architectures such as hypernetworks and online retraining to address issues like feature drift and variable feature sets.

Dynamic Feature Selection (DFS) is a machine learning setting in which a model does not assume that all features are always available or should always be used. Instead, it acquires, ranks, activates, or updates features dynamically—typically sequentially, per sample, per time step, or during training—under prediction, computation, communication, or acquisition-cost constraints. In the recent literature, the term most commonly denotes adaptive feature acquisition for partially observed inputs, where a policy chooses which feature to query next from the currently observed subset; however, the same term is also used for embedded sparse training, online feature updating under drift, and dynamic evaluation of feature-selection trajectories (Covert et al., 2023, Gadgil et al., 2023, Rajabinasab et al., 2024). The acronym is overloaded: in graph algorithms, “dynamic DFS” often refers instead to dynamic maintenance of depth-first-search trees, a distinct topic (Baswana et al., 2015).

1. Problem formulation and scope

A standard DFS formulation begins with a full input $x = (x_1,\dots,x_d)$ and label $y$, but assumes that only a subset of features has been observed at any decision point. For a currently observed subset $S \subseteq [d]$, the partial input is written $x_S = \{x_i : i \in S\}$. The system typically contains two learned components: a predictor $f(x_S;\theta)$, which must operate on arbitrary partial feature sets, and a selection or value function that scores candidate next features (Gadgil et al., 2023). The sequential policy then expands the subset one feature at a time, with the general objective of maintaining predictive quality while reducing acquisition cost or feature budget (Covert et al., 2023).

This setting differs fundamentally from static feature selection. Static methods choose one global subset and reuse it for every sample. DFS instead permits the selected subset to depend on the sample, the current uncertainty, the acquisition history, or the available budget. In the classification setting considered in several papers, the objective after acquiring $k$ features is expressed as

$$v_k(\pi,f) = \mathbb{E}_{p(x,y)} \left[ \ell\!\left(f\!\left(\{x_{i_t}\}_{t=1}^k\right),y\right) \right],$$

where the next queried feature depends on the previously acquired ones through a policy $\pi$ (Covert et al., 2023).

The same formal pattern is extended in multiple directions. Some work allows non-uniform feature costs, replacing pure information gain with information gain per unit cost; some permits variable stopping by introducing a stopping action $\pi(x_S)\in \{0\}\cup [d]$; and some conditions both prediction and acquisition on prior information $z$, yielding selection criteria of the form $y$0 (Gadgil et al., 2023). A further generalization relaxes the fixed feature universe assumption itself: in DFS with variable feature sets, each instance $y$1 has its own set $y$2, and the policy chooses among instance-specific candidate features using auxiliary descriptors $y$3 called “features of features” (Takahashi et al., 12 Mar 2025).

The literature therefore uses “dynamic” in several technically distinct senses. In the narrowest sense, DFS is sequential per-instance feature acquisition at inference time. In broader usage, it also includes time-varying reacquisition in multivariate monitoring, frame-by-frame re-ranking in tracking, drift-driven feature replacement in incremental learning, and feature selection embedded directly into the training dynamics of sparse or federated models. This suggests that DFS is best understood as an adaptive control layer over feature availability rather than as a single algorithmic family.

2. Information-theoretic and optimization foundations

A central idealization in modern DFS is the greedy conditional mutual information (CMI) rule

$y$4

which selects the next feature that is most informative about $y$5 given the already observed subset (Gadgil et al., 2023). The same rule appears in equivalent entropy-reduction form,

$y$6

and, for classification with cross-entropy, aligns with the Bayes predictor $y$7 (Covert et al., 2023). In regression, the corresponding criterion becomes reduction in conditional variance rather than reduction in conditional entropy (Gadgil et al., 2023).

The appeal of the CMI rule is theoretical clarity; its difficulty is implementation. Exact computation requires access to conditional distributions such as $y$8 and $y$9, which are generally unavailable. One line of work addresses this through amortized optimization: a policy network and predictor are trained jointly on one-step-ahead prediction loss so that, at the global optimum under expressive function classes, the learned policy recovers the greedy CMI solution (Covert et al., 2023). Another line avoids explicit generative modeling by learning a discriminative estimator of incremental utility. In that formulation, a value network is trained against the loss reduction

$S \subseteq [d]$0

and, if the predictor is Bayes-optimal with cross-entropy loss, the expectation of $S \subseteq [d]$1 equals $S \subseteq [d]$2 (Gadgil et al., 2023).

Recent work also challenges the shared-predictor assumption. A standard shared model minimizes expected loss over subsets,

$S \subseteq [d]$3

but the ideal subset-specific formulation assigns distinct parameters $S \subseteq [d]$4 to each subset $S \subseteq [d]$5 (Fumanal-Idocin et al., 12 May 2026). “Hypernetworks for Dynamic Feature Selection” proposes $S \subseteq [d]$6, where a hypernetwork $S \subseteq [d]$7 generates subset-specific classifier parameters conditioned on the observed mask, with Set Transformer encoding used to construct a smooth conditioning space over subsets (Fumanal-Idocin et al., 12 May 2026). The paper argues that this architecture addresses compromise across subsets, gradient conflicts, and sparse coverage of rare masks more directly than mask-concatenation methods.

A related extension appears when the feature universe is not fixed. In “Dynamic Feature Selection from Variable Feature Sets Using Features of Features,” the ideal policy becomes

$S \subseteq [d]$8

and the predictor becomes $S \subseteq [d]$9, so feature identity is represented explicitly through auxiliary descriptors $x_S = \{x_i : i \in S\}$0 rather than by coordinate position alone (Takahashi et al., 12 Mar 2025). This is a direct response to settings in which the measurable feature set varies across instances.

3. Sequential control, stopping, and partial-observation policies

Many DFS systems are formulated as sequential decision processes. In tabular prediction, a common training recipe pretrains a predictor on random masks, then jointly updates predictor and value network while features are selected by $x_S = \{x_i : i \in S\}$1, or by $x_S = \{x_i : i \in S\}$2 under non-uniform costs, with $x_S = \{x_i : i \in S\}$3-greedy exploration during training (Gadgil et al., 2023). This makes acquisition policy learning and partial-input prediction tightly coupled.

A more explicit control-theoretic formulation appears in constrained continuous data acquisition. There, the feature-selection process is modeled as a Markov Decision Process in which the state is the current partial subset $x_S = \{x_i : i \in S\}$4, actions are unselected features, and the reward is the sum of a machine-learning performance term and an exploration bonus. The performance term is based on F1-score computed from previously collected compatible data, while the exploration term is

$x_S = \{x_i : i \in S\}$5

and action selection follows an $x_S = \{x_i : i \in S\}$6-greedy rule (Sahin et al., 2020). The same work emphasizes delayed data arrival: the subset chosen at time $x_S = \{x_i : i \in S\}$7 yields data only at time $x_S = \{x_i : i \in S\}$8, so selection must be made before observing the current batch.

Clinical multivariate time-series introduce a more demanding sequential setting because features are not only selectively acquired but may need to be reacquired over time. In “Dynamic feature selection in medical predictive monitoring by reinforcement learning,” patient $x_S = \{x_i : i \in S\}$9’s data are represented as $f(x_S;\theta)$0, labels as $f(x_S;\theta)$1, and acquisition actions as a binary matrix $f(x_S;\theta)$2 of the same shape (Chen et al., 2024). The state carries forward old values for unselected features and refreshes selected ones, turning DFS into an online observation-policy problem. The policy is trained with PPO under a cost constraint, using a dynamic gate multiplier

$f(x_S;\theta)$3

to suppress cost reward as training cost approaches the budget (Chen et al., 2024). Because the acquisition policy induces a distribution shift from original sequences $f(x_S;\theta)$4 to masked states $f(x_S;\theta)$5, the predictor is subsequently retrained on synthetic states generated by replaying the learned policy.

Sequential control also arises in vision. For robust image matching, feature-layer selection is framed as an MDP in which each action either selects a CNN layer or terminates selection. The reward is

$f(x_S;\theta)$6

trading off compactness against final matching quality (Huang et al., 2020). The learned Q-network uses an encoder–BiLSTM–decoder architecture and is trained with Double Q-learning, target-network updates, a dueling architecture, and Retrace-style off-policy correction (Huang et al., 2020). A different but related online policy appears in object tracking, where the Dynamic Feature Selection Tracker re-ranks Color Names channels at each frame using a supervised adaptation of Infinite Feature Selection and then updates the ACT representation with the selected subset (Roffo et al., 2016).

Across these formulations, stopping is a first-class component rather than a post hoc threshold. Some methods use explicit terminate actions; others stop when the best information-gain-per-cost score falls below a penalty parameter $f(x_S;\theta)$7 (Gadgil et al., 2023). A plausible implication is that DFS is as much about adaptive stopping as about adaptive querying.

4. Embedded, online, and nonstationary variants

Not all DFS methods operate purely at inference time. Several systems embed feature selection into model training or update the feature space itself as the environment changes.

In personalized e-mail spam filtering, the nonstationary problem is twofold: concept drift changes the decision boundary, and feature drift changes which terms are discriminative. “Incremental personalized E-mail spam filter using novel TFDCR feature selection with dynamic feature update” addresses this by combining the TFDCR ranking function, incremental SVM learning, and a dynamic feature-update heuristic called selectionRankWeight (Sanghani et al., 2019). The current feature set $f(x_S;\theta)$8 is updated by adding new features whose normalized discriminative score exceeds the average score of the candidate set and removing old features with the lowest discriminative weight, while keeping dimensionality fixed: $f(x_S;\theta)$9 The paper notes that in each incremental pass, about $k$0–$k$1 of the existing feature set was updated (Sanghani et al., 2019).

In federated learning, DFS may be embedded directly into sparse optimization. “Embedded Federated Feature Selection with Dynamic Sparse Training: Balancing Accuracy-Cost Tradeoffs” introduces Dynamic Sparse Federated Feature Selection (DSFFS), where input-layer neurons, input connections, and hidden-layer connections are dynamically pruned and regrown during local client training (Mahanipour et al., 7 Apr 2025). Input feature importance is measured by the neuron strength

$k$2

and input neurons with the lowest strength are pruned, while disconnected neurons with the highest absolute gradient magnitudes are regrown (Mahanipour et al., 7 Apr 2025). In this formulation, feature selection is not a separate stage; removal of an input neuron is itself the feature-selection act.

Morphosyntactic analysis offers another notion of dynamicity: dynamic reordering during greedy feature-template selection. “Static and Dynamic Feature Selection in Morphosyntactic Analyzers” compares fixed human-designed template orderings with dynamic MRMR-style ordering based on mutual information and redundancy, recomputed after every accepted feature (Bohnet et al., 2016). The objective combines relevance and redundancy as $k$3, and the paper reports that feature selection can reduce runtime by up to $k$4 while maintaining or improving accuracy in some settings (Bohnet et al., 2016). Here “dynamic” refers not to per-instance acquisition, but to context-dependent re-ranking within the search over feature templates.

Unsupervised vision introduces yet another embedded form. “Unsupervised Dynamic Feature Selection for Robust Latent Spaces in Vision Tasks” proposes Dynamic Data Selection (DDS), an instance-wise selector that computes soft scores, converts them into hard top-$k$5 masks, and removes unselected pixels before they reach the downstream unsupervised model (Corcuera et al., 2 Oct 2025). The hard gating is designed so that irrelevant features are removed from the computation rather than merely down-weighted. The paper claims that DDS fills an open gap by providing the first unsupervised DFS approach that preserves $k$6D image structure and can be attached to arbitrary differentiable unsupervised architectures (Corcuera et al., 2 Oct 2025).

These variants show that DFS is not limited to budgeted tabular prediction. It also functions as a mechanism for drift adaptation, model sparsification, template redundancy control, and latent-space denoising.

5. Evaluation as a dynamic trajectory rather than a single-point score

A recurring difficulty in DFS research is evaluation. Standard downstream-task metrics such as classification accuracy or clustering accuracy assess only the final selected subset, while classical stability metrics such as Nogueira’s or Kuncheva’s assess overlap of selected feature identities across runs. “FSDEM: Feature Selection Dynamic Evaluation Metric” argues that both views are too narrow when the behavior of a method changes with subset size (Rajabinasab et al., 2024).

FSDEM evaluates a feature selector through its performance curve $k$7 over feature counts $k$8. The discrete observations are approximated by a continuous function $k$9, using linear interpolation, the trapezoidal rule for integration, and finite differences for derivatives (Rajabinasab et al., 2024). The main score is the normalized area under the performance curve,

$$v_k(\pi,f) = \mathbb{E}_{p(x,y)} \left[ \ell\!\left(f\!\left(\{x_{i_t}\}_{t=1}^k\right),y\right) \right],$$0

and the associated stability score is

$$v_k(\pi,f) = \mathbb{E}_{p(x,y)} \left[ \ell\!\left(f\!\left(\{x_{i_t}\}_{t=1}^k\right),y\right) \right],$$1

If the underlying metric lies in $$v_k(\pi,f) = \mathbb{E}_{p(x,y)} \left[ \ell\!\left(f\!\left(\{x_{i_t}\}_{t=1}^k\right),y\right) \right],$$2, FSDEM is bounded accordingly, and higher values are better (Rajabinasab et al., 2024).

This stability notion is not subset-overlap stability. It measures how smoothly or consistently performance changes as features are added. The distinction is operationally important in the paper’s dummy example with semantically redundant representations such as salary in EUR versus salary in USD, or distance in kilometers versus distance in miles: exact-overlap metrics regard such selections as unstable, while FSDEM’s derivative-based score treats them as stable if performance remains essentially unchanged (Rajabinasab et al., 2024). This suggests that, for DFS, “stability” can refer to information-equivalent acquisition behavior rather than literal identity preservation.

The empirical setup in FSDEM uses 20 UCI datasets and compares random feature selection, information gain, chi-square, and a wrapper-based random forest method under metrics including classification accuracy and clustering accuracy (CLACC) (Rajabinasab et al., 2024). Among the reported observations are that the best method can depend on the target subset-size range, that wrapper-based random forest tends to have the best overall FSDEM score, and that approximating FSDEM with half the observations produces only a small difference (Rajabinasab et al., 2024). For stage-wise or incremental selectors, this trajectory-based evaluation is directly aligned with the underlying process being modeled.

6. Interpretability, uncertainty, and persistent challenges

A common criticism of modern DFS is that many strong methods are opaque. “Dynamic Feature Selection based on Rule-based Learning for Explainable Classification with Uncertainty Quantification” addresses this by using a rule-based global model and defining DFS as the process of recovering the global model’s prediction from a partial subset (Fumanal-Idocin et al., 4 Aug 2025). The aleatoric component of feature value is measured by the divergence between full-model and partial-model predictions,

$$v_k(\pi,f) = \mathbb{E}_{p(x,y)} \left[ \ell\!\left(f\!\left(\{x_{i_t}\}_{t=1}^k\right),y\right) \right],$$3

while epistemic uncertainty is measured by disagreement among active rules,

$$v_k(\pi,f) = \mathbb{E}_{p(x,y)} \left[ \ell\!\left(f\!\left(\{x_{i_t}\}_{t=1}^k\right),y\right) \right],$$4

Candidate features are then scored through

$$v_k(\pi,f) = \mathbb{E}_{p(x,y)} \left[ \ell\!\left(f\!\left(\{x_{i_t}\}_{t=1}^k\right),y\right) \right],$$5

and the rule structure is used to restrict the active search space to features appearing in still-viable rules (Fumanal-Idocin et al., 4 Aug 2025). The paper also shows that, under a proportionality condition, minimizing expected aleatoric uncertainty is equivalent to maximizing conditional mutual information (Fumanal-Idocin et al., 4 Aug 2025).

Interpretability concerns intersect with several other open issues. One is cost heterogeneity: some methods explicitly support non-uniform costs and gain-per-cost policies, while others note support for heterogeneous costs as future work (Gadgil et al., 2023, Fumanal-Idocin et al., 4 Aug 2025). Another is distribution shift induced by the policy itself, which motivates predictor retraining on policy-generated states in medical monitoring (Chen et al., 2024). A third is generalization to unseen acquisition patterns; the hypernetwork literature treats this as a structural problem and reports stronger zero-shot generalization to feature subsets never seen during training than mask-conditioned shared predictors (Fumanal-Idocin et al., 12 May 2026). A fourth is variation in the measurable feature universe, addressed by conditioning on “features of features” rather than relying on a fixed coordinate system (Takahashi et al., 12 Mar 2025).

The literature also contains persistent misconceptions. One is to equate DFS with static subset ranking followed by truncation; the sequential formulations in tabular, medical, image, and sensing settings all make the next choice depend on the currently observed subset (Covert et al., 2023, Sahin et al., 2020). Another is to treat stability solely as identity overlap, even when informational equivalence is the relevant object (Rajabinasab et al., 2024). A third is to assume that a single predictor conditioned on a mask is always sufficient; recent hypernetwork work argues that the exponential family of possible observation patterns can require subset-specific specialization (Fumanal-Idocin et al., 12 May 2026).

Taken together, these developments place DFS at the intersection of active feature acquisition, uncertainty-aware decision-making, sparse learning, and evaluation methodology. The field now spans greedy information-theoretic policies, discriminative value regression, reinforcement learning under delayed or recurrent observation, dynamic sparse training in federated systems, unsupervised latent-space filtering, and explainable rule-based querying. The common denominator is not a single optimizer or architecture, but the premise that feature relevance is contingent on state, budget, and context rather than fixed once and for all.