Sensitivity Score for Unlabeled Nodes

• The paper introduces formal definitions and computational techniques for estimating sensitivity scores in unlabeled nodes across classifiers and network structures.
• It employs algorithmic methods like EM/Bayesian-MCMC, Fréchet derivatives, and semi-supervised deep learning to quantify node impact without gold-standard labels.
• Practical guidelines are provided for adapting sensitivity estimation approaches to network structure, uncertainty quantification, and computational scalability.

A sensitivity score for unlabeled nodes quantifies the latent impact, uncertainty, or influence of a node within a network or for a classifier, in settings where ground-truth labels are absent or only partially available. The concept underlies both the evaluation of models (such as classifier sensitivity without gold-standard labels) and the structural or functional assessment of nodes in networks (such as communicability sensitivity or importance under uncertainty). This article surveys the formal definitions, estimation techniques, algorithmic methodologies, and domain-specific implementations of sensitivity scores for unlabeled nodes, drawing on recent literature spanning classification diagnostics, active network sampling, importance estimation with uncertainty, and matrix-function network analysis.

1. Formal Definitions and Interpretations

A sensitivity score is context-dependent and admits multiple rigorous definitions. In binary classification without labeled data, sensitivity is the true positive rate, $Se = P(\hat Y=1\,|\,Y=1)$, estimated when $Y$ is unobserved (Evans, 2022). In network analysis, it quantifies the effect of removing or modifying a node on communicability or related graph indices (Schweitzer, 2023). For semi-supervised importance estimation, sensitivity is modeled as the predictive uncertainty (variance) for a node's importance score (Chen et al., 26 Mar 2025). In active learning for unlabeled graphs, node-specific utility or centrality scores serve as surrogates for "sensitivity," indicating which nodes' labels would most reduce downstream classification error (Kajdanowicz et al., 2015).

Distinguishing variants include:

• Diagnostic Sensitivity for classifiers: $Se$ as the fraction of positives predicted positive, estimated without labeled outcomes (Evans, 2022).
• Communicability Sensitivity: The Fréchet derivative of a network functional (e.g., $C^{\mathrm{TC}}$ for total communicability) under node removal (Schweitzer, 2023).
• Uncertainty-based Sensitivity: The standard deviation of a modeled importance distribution $s_x \sim \mathcal N(\mu, \sigma^2)$, i.e., $\sigma$, for node $x$ (Chen et al., 26 Mar 2025).

2. Sensitivity Estimation Without Node Labels

Estimating sensitivity without access to true node labels necessitates structural, probabilistic, or surrogate approaches.

• In classifier diagnostics, the empirical marginal positive rate $q = P(\hat Y=1)$ is linked to sensitivity, prevalence, and specificity via $q = \pi Se + (1-\pi)(1-Sp)$, introducing a nonlinear constraint system with latent class proportions and error rates (Evans, 2022).
• In positive-unlabeled (PU) learning, estimation leverages a labeled subset of positives $P$ and an unlabeled set $U$, expressing sensitivity or TPR at threshold $r$ using the number of positives in $P$ and the (unknown) proportion of positives $\pi_u$ in $U$: $$\widehat{\rm TPR}(r) = \frac{a + \pi_u b}{m + \pi_u n}$$, with $a=|P \cap \text{top-}r|$, $b=|U \cap \text{top-}r|$ (Claesen et al., 2015).

In both cases, identifiability typically hinges on (1) access to multiple independent predictions (e.g., different classifiers), (2) variation in class prevalence across subpopulations, or (3) informative priors or bounding constraints on prevalence or performance. When such additional information is unavailable, only partial or interval estimates are attainable.

3. Structural and Centrality-Based Sensitivity Scores

For networked, unlabeled data, sensitivity or utility scores exploit structural measures. Seven classic centrality functions formalize utility, each computable from the graph topology without labels (Kajdanowicz et al., 2015):

• In-degree/out-degree centrality: $U_{\rm in}(v) = d^{\rm in}(v)$, $U_{\rm out}(v) = d^{\rm out}(v)$.
• Betweenness: Summed fraction of shortest paths through node $v$.
• Clustering coefficient: Fraction of neighbor pairs of $v$ that are interconnected.
• HITS hub/authority: Recursive high-importance nodes.
• PageRank: Stationary probability from a random walk.

These scores can be directly used to rank unlabeled nodes according to their prospective informativeness for label acquisition or model retraining. "Measure–neighbour" strategies further refine sensitivity scoring by querying neighbors of structurally extreme nodes, which improves label propagation in high-clustering networks.

4. Algorithmic and Statistical Methods for Sensitivity Computation

Several algorithmic frameworks operationalize sensitivity estimation for unlabeled nodes:

• EM/Bayesian-MCMC for Classifier Metrics: Expectation-Maximization and MAP-EM iterations jointly estimate latent prevalence, sensitivity, and specificity, updating parameters via posterior expectations. Beta priors enforce regularization (Evans, 2022).
• Closed-form and Bootstrap Intervals for PU Setting: The point estimator for sensitivity is augmented with DKW-based confidence bounds, contingent on the representativeness of known positives (Claesen et al., 2015).
• Low-Rank Matrix Function Sensitivities: The Fréchet derivative $L_{\exp}(A, E_v)$, where $E_v$ encodes node removal, quantifies node-sensitivity for total communicability, subgraph centrality, and the Estrada index. Efficient two-sided Krylov approximations enable computation of all $n$ node sensitivities at $\mathcal O(n m)$ cost for sparse $A$, with small $m$ (Schweitzer, 2023).
• Semi-supervised Deep Learning (EASING/DJE): The Distribution-based Joint Estimator (DJE) embeds each node, then outputs a Gaussian node-importance distribution, with the variance (or log-variance) serving as the sensitivity score. Monte Carlo dropout and ensembling are used for pseudo-labeling; a heteroscedastic loss penalizes unreliable pseudo-labels (Chen et al., 26 Mar 2025).

The following table contrasts representative sensitivity score types and algorithms:

Context	Sensitivity Score	Algorithmic Approach
Classifier	$Se$ (TPR)	EM, closed-form, Bayesian
PU Learning	$\widehat{\rm TPR}$	Contingency-table + prevalence est.
Network	Centrality/Utility	Degree, betweenness, clustering
Communicability	$S^\mathrm{TC}_v$	Fréchet derivative, Krylov
Semi-supervised importance	$\sigma_x$	DJE uncertainty, MC dropout

5. Sensitivity Under Uncertainty and Practical Guidelines

Sensitivity estimation in unlabeled settings is confounded by model or data uncertainty, model identifiability, and structural features:

• Uncertainty Quantification: In semi-supervised and distributional setups, node-level variance quantifies sensitivity, and the loss is weighted to down-tune contributions from high-uncertainty nodes, thus regularizing over-fitting to unreliable pseudo-labels (Chen et al., 26 Mar 2025).
• Network Structure Effects: The efficacy of different sensitivity-based sampling or seeding strategies depends strongly on network clustering and modularity. Measure-based strategies are preferred for low clustering, while measure–neighbour and LBP-based strategies excel for high clustering (Kajdanowicz et al., 2015).
• Computational Complexity: For matrix-function sensitivities, low-rank Fréchet–Krylov methods achieve scalability by reusing subspaces and exploiting the sparsity of $A$, enabling sensitivity evaluation on networks of millions of nodes (Schweitzer, 2023).

6. Theoretical Bounds, Validity, and Limitations

A priori bounds on sensitivity scores are available in the matrix-function context based on graph distances and polynomial approximation theory. For instance, the magnitude of $[L_{\exp}(A,E_{ij})]_{u,v}$ decays super-exponentially in terms of $d(u,i) + d(j,v)$, allowing screening-type approximations to identify only locally sensitive nodes (Schweitzer, 2023). Identifiability and estimator calibration in classifier diagnostics require conditional independence, prevalence shifts, or strong priors; violations bias sensitivity estimates and require empirical sensitivity analysis (Evans, 2022, Claesen et al., 2015).

Common limitations include: under-identifiability in single-dataset scenarios, reliance on the representativeness of partial labels, and possible instability in highly disconnected or poorly observed network regions. Confidence intervals derived for sensitivity statistics relate to sampling noise but do not account for possible model misspecification.

7. Applications and Recommendations

Sensitivity scores for unlabeled nodes are applicable to classifier validation without gold standards (bioinformatics, medical testing), active node sampling in collective classification, node-importance modeling under uncertainty, and robustness analysis in complex networks. Method selection is context-dependent:

• For classifier evaluation, use EM or Bayesian estimation with informative priors and assess identifiability.
• For active node selection, choose seeding strategies based on structured utility scores, adapting to network clustering and connectivity.
• For scalable network analysis, leverage low-rank matrix-function derivative techniques.
• For uncertainty-regularized node ranking, employ semi-supervised methods like DJE to produce both importance and sensitivity for unlabeled nodes.

Robust application mandates careful checking of structural or probabilistic assumptions, cross-validation against random baselines, and quantification of estimator variance, supported by model-driven or data-driven sensitivity analyses (Evans, 2022, Kajdanowicz et al., 2015, Chen et al., 26 Mar 2025, Schweitzer, 2023, Claesen et al., 2015).