Locally Adaptive Scoring (ARES)

• Locally Adaptive Scoring (ARES) is a family of techniques that adjust scoring functions based on local structural, statistical, or semantic properties.
• These methods integrate locality-aware components like nearest-neighbor reconstruction and kernel smoothing to improve anomaly detection and conformal prediction accuracy.
• Empirical results show ARES enhances performance in detecting anomalies and risk estimation, offering greater robustness in heterogeneous and high-dimensional domains.

Locally Adaptive Scoring (ARES) encompasses a family of methods characterized by data-driven, context-sensitive adaptation of scoring or risk estimation functions to local structural, statistical, or semantic properties of the input space. Prominent instantiations across anomaly detection, conformal regression, and healthcare risk estimation exhibit foundational principles of local adaptivity, typically leveraging locality structures such as neighborhoods, kernel smoothing, or self-attention to robustly adjust predictions or scores in high-dimensional, heterogeneous domains.

1. Principles and Motivation

Locally adaptive scoring methods address inherent limitations of global scoring rules that neglect spatial or predictive heterogeneity in high-dimensional data distributions. Standard anomaly detectors, conformal inference intervals, and risk scores often fail to adapt to local variations in reconstruction error, conditional uncertainty, or risk profile, resulting in suboptimal detection, poorly calibrated intervals, or inaccurate risk estimates (Goodge et al., 2022, Deutschmann et al., 2023, Renc et al., 10 Feb 2025). ARES variants explicitly model or estimate locality-dependent statistics—such as conditional error medians, local densities, or trajectory-dependent risk probabilities—leading to more robust and interpretable inferences.

2. Adaptive Reconstruction Error-Based Scoring

The ARES method in anomaly detection, introduced in "ARES: Locally Adaptive Reconstruction-based Anomaly Scoring" (Goodge et al., 2022), enhances autoencoder-based anomaly scoring by incorporating local adaptivity in the latent space of normal training data. The method leverages two locality-aware components: (i) a local reconstruction score and (ii) a local density score.

Let $x \in \mathbb{R}^d$ denote an input sample, $g_\varphi(x) = z \in \mathbb{R}^k$ its latent representation with encoder parameters $\varphi$, and $f_\theta(z) = \hat{x}$ its reconstruction with decoder parameters $\theta$. The reconstruction residual is $\epsilon = x - \hat{x}$, with associated error $R(\epsilon) = \|\epsilon\|_2^2$. For a training set of $m$ normal samples, their latent encodings $\{z_i\}_{i=1}^m$ and reconstruction errors $\{R_i\}_{i=1}^m$ are stored.

For a test query $g_\varphi(x) = z \in \mathbb{R}^k$0:

• The $g_\varphi(x) = z \in \mathbb{R}^k$1-nearest latent neighbors $g_\varphi(x) = z \in \mathbb{R}^k$2 are identified under Euclidean distance.
• The local reconstruction score is defined as

$g_\varphi(x) = z \in \mathbb{R}^k$3

• The local density score is approximated as $g_\varphi(x) = z \in \mathbb{R}^k$4, with $g_\varphi(x) = z \in \mathbb{R}^k$5 denoting the k-local outlier factor, or via alternative nonparametric density estimators.

The final ARES anomaly score is:

$g_\varphi(x) = z \in \mathbb{R}^k$6

with $g_\varphi(x) = z \in \mathbb{R}^k$7 (default $g_\varphi(x) = z \in \mathbb{R}^k$8) controlling the balance between local error deviation and local sparsity.

Significance emerges from ARES’s ability to detect samples that are anomalous within their local neighborhood, thus exhibiting robustness to naturally heteroscedastic or multimodal well-posed regions in the data. Empirical results across benchmark datasets (MNIST, FMNIST, SNSR, OTTO, MI-F) show consistent Area-Under-ROC (AUC) improvements over baseline autoencoder scoring by up to +18.3 AUC points in highly heteroscedastic settings (Goodge et al., 2022).

3. Locally Adaptive Conformal Prediction

In regression and predictive inference, the ARES methodology of "Adaptive Conformal Regression with Jackknife+ Rescaled Scores" (Deutschmann et al., 2023) addresses the limitation of conformal prediction intervals that provide only global coverage guarantees, neglecting spatially varying error structures. The core technical advance is the integration of local rescaling of calibration scores with a Jackknife$g_\varphi(x) = z \in \mathbb{R}^k$9 conformal construction, preserving exchangeability and global coverage while increasing local adaptivity.

Given i.i.d. calibration samples $\varphi$0, a point predictor $\varphi$1 is fit, and nonconformity scores $\varphi$2 (e.g., $\varphi$3) are calculated. Rather than directly thresholding the global ECDF of $\varphi$4, the ARES method rescales each score by a locally estimated conditional mean:

$\varphi$5

with the local mean estimated via Nadaraya–Watson kernel smoothing:

$\varphi$6

A Jackknife$\varphi$7 scheme, using leave-two-out rescaling, ensures exchangeability and preserves global coverage:

$\varphi$8

Empirically, this procedure substantially improves finite-sample local coverage and yields tighter intervals in heteroscedastic or low-sample regimes. The method is robust on molecular regression, antibody-binding affinity, and vision benchmarks, outperforming ECDF reweighting approaches in many sub-asymptotic scenarios (Deutschmann et al., 2023).

4. Foundation Models for Adaptive Risk Estimation

In predictive healthcare, the Adaptive Risk Estimation System (ARES) (Renc et al., 10 Feb 2025) provides dynamic, personalized risk probabilities for clinical outcomes by leveraging the ETHOS foundation model—a transformer-based generative model for tokenized electronic health record (EHR) sequences. Here, locality and adaptivity are realized via both Monte Carlo simulation of possible patient health trajectories (PHTs) and attention-based attribution mechanisms.

Risk for event $\varphi$9 is computed by generating $f_\theta(z) = \hat{x}$0 future token sequences $f_\theta(z) = \hat{x}$1 from a PHT up to time $f_\theta(z) = \hat{x}$2, with the empirical probability estimator:

$f_\theta(z) = \hat{x}$3

where $f_\theta(z) = \hat{x}$4 if $f_\theta(z) = \hat{x}$5 occurs in $f_\theta(z) = \hat{x}$6 during simulation.

Local adaptivity arises from:

• Self-attention mechanisms: ETHOS’s attention weights $f_\theta(z) = \hat{x}$7 dynamically learn which historical clinical events or codes are most prognostically relevant for each patient at each time step.
• Personalization: Risk estimates are conditioned on the entire patient timeline, demographic tokens, time-encoded features, and comorbidities, with no static thresholds or feature selection.
• Explanatory feedback: Attribution scores $f_\theta(z) = \hat{x}$8 highlight influential input events on risk scores.

Evaluation on MIMIC-IV data demonstrates that ARES (ETHOS-based) outperforms traditional early warning scores and ML comparators (e.g., NEWS, logistic regression, Med2Vec) in AUROC across multiple critical outcomes, with robust calibration and strong subgroup performance (Renc et al., 10 Feb 2025).

5. Implementation and Computational Considerations

The computational overhead of locally adaptive scoring is largely determined by locality-defining operations:

• Nearest-neighbor searches in latent or input spaces (used in anomaly detection and density estimation) can be optimized via KD-tree or HNSW indexing, reducing test-time cost from $f_\theta(z) = \hat{x}$9 (exact) to $\theta$0 (approximate) (Goodge et al., 2022).
• Kernel-based local scale estimation in conformal regression requires $\theta$1 kernel evaluations but can be made tractable with sparse KNN graphs or random feature projections (Deutschmann et al., 2023).
• In generative sequence models, the principal cost is in autoregressive sampling and attention computation, but inference times remain compatible with clinical real-time feedback needs (seconds per instance) (Renc et al., 10 Feb 2025).

Hyperparameter selection (e.g., neighborhood size $\theta$2, kernel parameters, mixture weights $\theta$3) is data-dependent. Robustness arises when local error distributions are unimodal and the density estimator is well-calibrated to sample sparsity.

6. Empirical Results and Comparative Performance

Across domains, locally adaptive ARES-style methods consistently outperform globally fixed baselines:

Domain	Task/Metric	Baseline	ARES	Δ Performance
Anomaly Detection (Goodge et al., 2022)	MNIST (one-class, AUC %)	AE: 96.96	97.89	+0.93
	OTTO (one-class, AUC %)	AE: 85.26	87.86	+2.60
	MI-F (one-class, AUC %)	AE: 71.19	89.52	+18.33
Conformal Regression (Deutschmann et al., 2023)	Antibody affinity (interval size, low $\theta$4)	LVD	ARES	Smaller/inter.
Healthcare Risk (Renc et al., 10 Feb 2025)	Hospital mortality (AUROC)	MEDS-Tab: 0.887	ETHOS-ARES: 0.936	+0.049
	ICU admission (AUROC)	MEDS-Tab: 0.918	ETHOS-ARES: 0.927	+0.009

Performance gains are most pronounced in regimes with high heteroscedasticity, multimodality, or low calibration sample size.

7. Limitations and Open Directions

Despite their empirical successes, ARES methodologies possess domain-specific limitations:

• Local adaptivity may be undermined by contaminated or inadequately clustered training data, particularly in autoencoder-driven latent spaces (Goodge et al., 2022).
• Kernel and density estimator choice in conformal or anomaly settings may require tuning for stability across highly heterogeneous or high-dimensional inputs, and the method may not optimally capture multimodal conditional error distributions (Deutschmann et al., 2023).
• Generalizability of foundation model-based local adaptivity (e.g., ETHOS-ARES) to out-of-distribution hospital systems remains a challenge due to domain shift and coding heterogeneity; additional data modalities and causal trajectory analyses remain open research topics (Renc et al., 10 Feb 2025).
• Theoretical coverage guarantees in locally adaptive conformal methods are currently marginal over calibration randomness; strong PAC-style guarantees conditional on calibration are an open problem (Deutschmann et al., 2023).

A plausible implication is that ongoing work in scalable local statistics estimation, causal sequence modeling, and interpretable attention mechanisms will further advance the reliability and applicability of locally adaptive scoring systems across settings demanding both local fidelity and global guarantees.