Feature Utility Metric

• Feature utility metric is a quantitative measure that evaluates the usefulness, stability, and informativeness of features with respect to model objectives.
• It employs methodologies like RoSHAP, AGF, and FSDEM to enhance the reliability of feature selection and model pruning.
• This metric is applied across diverse domains, including deep learning, biometrics, and quantum computing, to improve robustness and practical performance.

A feature utility metric is a quantitative criterion that measures the usefulness, stability, or informativeness of features, feature sets, or feature representations with respect to model objectives or downstream tasks. These metrics are foundational in model interpretation, feature selection, structured pruning, robust attributions, quantum feature design, and applied perception tasks. Contemporary research develops domain-specific and theoretically principled metrics that go well beyond simple raw importance scores, emphasizing robustness to stochastic or adversarial variation, context-specific task utility, and multidimensional evaluation.

1. Foundations: Definitions and Purposes

Attribution-based feature utility metrics emerged to summarize individual feature contributions to model predictions post hoc, often via importance scores from methods such as SHAP or gradients. More generally, feature utility quantifies the practical or theoretical value a feature (or group of features) delivers to some aspect of learning or inference. The purposes include:

• Reliable feature selection and screening (identifying “signal” vs. “noise” (Xiang et al., 14 May 2026))
• Structural pruning and compression (preserving functional pathways or expressive capacity (Qian et al., 12 Mar 2026))
• Measuring informativeness and redundancy in feature maps (e.g., quantum embeddings (Illésová et al., 30 Jun 2025))
• Guiding robust interpretations for high-stakes domains (e.g., genomics, face biometrics (Babnik et al., 29 Apr 2026))

A robust feature utility metric must remain stable under variation in data sampling, model initializations, or underlying optimization procedures, and must penalize features whose apparent utility is highly stochastic.

2. Distributional and Robust Feature Utility Metrics

Classical single-run feature attribution metrics are vulnerable to instability: different data splits or seeds can yield irreproducible or misleading rankings. The robust SHAP-based feature utility metric, RoSHAP, addresses this by modeling attribution as a stochastic process and aggregating over its empirical distribution (Xiang et al., 14 May 2026).

Given SHAP attributions $T_{ij}$ for feature $j$ on sample $i$, the model posits $T_{ij}$ follows a zero-inflated mixture: $T_{ij}\sim w_{ij}\delta_0 + (1-w_{ij})G_{ij}$, with $B_{ij}=\mathbf{1}\{T_{ij}=0\}$ and $H_{ij}=|G_{ij}|$. The aggregated importance $U_j=\sum_{i=1}^n |T_{ij}|$ is treated as a random variable. RoSHAP summarizes the distribution with three quantities:

• Activity: $1-P_{0j}$, the empirical probability that $U_j$ is nonzero across resamples
• Strength: $j$0, the median of nonzero $j$1
• Stability: $j$2, the standard deviation of $j$3

The unified utility is $j$4, simultaneously rewarding features that are consistently active, strong in effect, and stable across draws. RoSHAP is computed via bootstrap resampling and kernel density estimation (or Gaussian approximation for large $j$5), and has well-defined asymptotic properties via the Lyapunov central limit theorem. Empirically, RoSHAP outperforms single-run scores in identifying true signal features and yields parsimonious models with high accuracy (Xiang et al., 14 May 2026).

3. Structural and Kinetic Utility in Deep Networks

Traditional structural pruning metrics in deep learning, such as weight $j$6-norm or activation magnitude, are susceptible to magnitude bias: they risk pruning small-but-functional channels. The Alternating Gradient Flow (AGF) utility metric circumvents this by quantifying each feature's “kinetic utility” via its cumulative gradient-driven loss variation along a calibration trajectory (Qian et al., 12 Mar 2026):

$j$7

This metric integrates absolute feature-space gradients, capturing both response amplitude and sensitivity, and is especially robust at extreme sparsity, preserving subnet topologies that achieve high post-pruning accuracy. AGF’s gradient–signal decoupling prevents signal cancellation and outperforms magnitude-based baselines—particularly under high compression or in architectures like Vision Transformers where attention and routing induce nontrivial sparsity bottlenecks (Qian et al., 12 Mar 2026).

In practice, AGF supports hybrid systems where expensive offline calibration (structural search) using AGF is combined with efficient, zero-cost online routing or confidence priors for dynamic inference.

4. Utility Metrics for Feature Selection and Evaluation

Expressive evaluation of feature selection algorithms requires metrics that capture both performance and stability across feature subset sizes. Standard downstream-only metrics (accuracy, F1) and set-overlap stability indices (e.g., Kuncheva, Nogueira) fail to encapsulate these. The Feature Selection Dynamic Evaluation Metric (FSDEM) provides two joint metrics (Rajabinasab et al., 2024):

• FSDEM Score:

$j$8

where $j$9 interpolates observed performance $i$0 as a function of feature count $i$1.

• Stability Score:

$i$2

with $i$3 the finite-difference derivative.

FSDEM integrates performance over the full range $i$4 of subset sizes. Stability captures the average marginal gain/loss when adding features. This pair of metrics enables sensitive, sample-efficient, and overlap-invariant evaluation, robust even when features are highly redundant (Rajabinasab et al., 2024).

Empirical results demonstrate that methods like random forest wrappers attain high FSDEM (0.85–0.95), while random selection inflates stability ($i$50.3–0.6) simply by monotonic feature addition.

5. Domain-Specific Utility Metrics: Biometrics, Robotics, Quantum

Feature utility is operationalized differently across domains, adapting to context-specific signal and noise structures.

Face Recognition and Biometric Utility

In face recognition, biometric utility is increasingly optimized via sample-specific measures linked to recognition outcomes, not classic image “quality” proxies. The Certainty Ratio (CR) (Babnik et al., 29 Apr 2026):

$i$6

reflects a sample's closeness to its own class center and separation from background classes in the embedding space. Incorporated into the FunFace adaptive margin loss as a normalized mix with feature norm, CR enables improved recognition—particularly in low-quality surveillance and drone data—over models relying only on visual quality.

Geometric Feature Metrics in Robotic Perception

In vision-based localization, the Geometric Feature Metric (GFM) guides planning by quantifying the informativeness of the environment at candidate robot poses (Lin et al., 22 Jul 2025):

$i$7

where $i$8 is the Jacobian (w.r.t. pose) for the $i$9-th LiDAR ray. Lower $T_{ij}$0 indicates locally richer geometries and better localizability. GFM is precomputed as a metric encoding map for real-time use in A*-search and trajectory optimization, minimizing localization error by proactively routing through feature-rich regions.

Quantum Feature Utility

For quantum models, QMetric quantifies the information-theoretic, geometric, and statistical value of quantum feature maps (Illésová et al., 30 Jun 2025):

• Feature Map Compression Ratio (FMCR): measures variance compression
• Effective Dimension (EDQFS): smooth estimate of latent dimensionality
• Quantum Layer Activation Diversity (QLAD): normalized entropy of post-encoding activations
• Quantum Output Sensitivity (QOS): finite-difference robustness metric

In practice, combinations of high FMCR, low EDQFS/QLAD, and high QOS identify pathological “collapsed” feature maps despite overall circuit expressibility or fidelity. The metrics provide actionable diagnosis for quantum circuit engineering.

6. Comparative Analysis and Theoretical Properties

Recent metrics offer several advantages over traditional approaches:

Metric Type	Captures Stability	Context Sensitivity	Penalizes Redundancy/Collapse	Empirical Robustness
Single-run attribution	No	Low	No	Low
RoSHAP (Xiang et al., 14 May 2026)	Yes	High (distributional)	Yes	High
AGF (Qian et al., 12 Mar 2026)	Yes	High (topological effect)	Yes	High
FSDEM (Rajabinasab et al., 2024)	Yes	High (performance curve)	Yes	High
Certainty Ratio (Babnik et al., 29 Apr 2026)	Yes	Yes (biometric context)	N/A	High
QMetric (Illésová et al., 30 Jun 2025)	Yes	Yes (activation/geometric context)	Yes	High

Theoretical properties such as asymptotic normality (RoSHAP), implicit regularization (AGF), and robustness to feature redundancy (FSDEM) underwrite these advantages. Most are accompanied by practical algorithms that scale to large models and datasets.

7. Implementation and Practical Guidelines

• For RoSHAP, select $T_{ij}$1 according to convergence; use Gaussian approximation for large $T_{ij}$2, otherwise $T_{ij}$3–$T_{ij}$4 bootstraps. Visualize $T_{ij}$5 distributions.
• For AGF, calibrate on a short training trajectory (even $T_{ij}$620 batches can suffice) before pruning; perform all routing or early exit decisions with static priors to avoid runtime overhead.
• For FSDEM, sample $T_{ij}$7 at a suitable grid; 50% sparsity in $T_{ij}$8-values changes results negligibly.
• For biometric utility, estimate Certainty Ratio per sample within the forward pass, and mix with norm-based scores in the adaptive loss.
• For GFM, precompute the metric encoding map and integrate tightly with trajectory planners (A*, MINCO) for real-time usage.
• For QMetric, leverage the API to compute all four metrics post-training; abnormal FMCR, QLAD, EDQFS, or QOS collectively diagnose major representational failure modes.

A consistent principle is that feature utility metrics are maximally informative only when integrated with robust resampling, performance curve analysis, and context-aware evaluation, as established by recent advances across modalities and domains.