Catching Disguised Transients with ASTRANet: Anomaly-Aware Spectroscopic Classification and Conformal Calibration

Technical Summary: Catching Disguised Transients with ASTRANet

Problem Statement
Time-domain surveys (e.g., ZTF, future LSST) generate hundreds of thousands of transient candidates annually, yet spectroscopic follow-up is severely rate-limited. Current deep learning classifiers for spectroscopic transient identification are typically "closed-set," meaning they force every input spectrum into a known training class with a softmax probability, regardless of whether the spectrum genuinely belongs to that class. This architecture creates a dangerous failure mode: high-confidence wrong answers. These classifiers often fail to identify rare, physically peculiar, or novel objects (anomalies) because they are trained to minimize error on common classes, effectively concealing the very failures that lead to discovery. Existing remedies often rely on single-score anomaly detection or photometric light curves, treating uncertainty quantification (UQ) and anomaly detection (AD) as separate axes, which the authors argue is insufficient for the physically diverse failure modes of spectroscopic data.

Methodology: The ASTRANet Framework
The authors present ASTRANet, a confidence-aware framework designed for spectroscopic transient classification and discovery. It integrates three coupled modules:

1. Hierarchical Spectral Classifier (ASTRANet):

   • Input: Operates directly on observer-frame spectra without requiring host-galaxy redshift or spectral phase as mandatory inputs. These are treated as optional metadata injected via Feature-wise Linear Modulation (FiLM) and randomly masked during training to ensure the network learns redshift-invariant representations.
   • Architecture: A deep ensemble of six models built on a multi-scale convolutional stem and a dilated temporal convolutional backbone (TCN). It employs multi-head attention pooling to generate a 192-dimensional embedding.
   • Output: Jointly predicts a coarse superclass (Thermonuclear, Core-Collapse, Non-SN) and a fine-grained class (7 classes: Ia, Ibc, SLSN, II, AGN, CV, TDE). The fine head consumes the coarse softmax as a soft prior via a stop-gradient mechanism.

2. Anomaly Detection Layer (ASTRANet-Sentinel):

   • Mechanism: Computes 16 complementary anomaly scores from the ensemble embeddings and logits, spanning four physically motivated families:
     • Uncertainty: (e.g., predictive entropy, ODIN, energy).
     • Distance: (e.g., Mahalanobis distances, kNN, cosine distance).
     • Density: (e.g., GMM likelihoods, isolation forest, LOF).
     • Hybrid: (e.g., hierarchical entropy combination, manifold residual scores).
   • Integration: These scores are non-linearly combined using a gradient-boosted decision tree (LightGBM) into a single cross-family anomaly score ($s_{AD}$). The authors demonstrate that a non-linear combiner is essential because no single family of scores can detect all failure modes.
   • Training: The encoder undergoes "outlier-exposure" fine-tuning, where an entropy term and embedding-margin term push exposure anomalies toward a uniform distribution and away from class centroids, significantly improving recovery of near-out-of-distribution (OOD) objects.

3. Conformal Uncertainty Quantification Layer (ASTRANet-CP):

   • Calibration: Uses Mondrian conformal prediction to calibrate the anomaly scores and classifier outputs. This provides distribution-free, finite-sample guarantees valid under arbitrary distribution shifts.
   • Outputs: For each spectrum, it generates:
     • A calibrated Mondrian p-value ($p_M$) indicating the probability the prediction is mistaken.
     • An adaptive prediction set (APS) of plausible classes at a user-specified coverage level.
     • Per-class ensemble probability intervals ($\bar{p}_k \pm \sigma_k$) capturing deep-ensemble disagreement.
     • A decomposition of predictive entropy into aleatoric (data noise) and epistemic (model disagreement) components.
   • Innovation: Introduces AD-stratified Mondrian Conformal Prediction (AD-MCP), which stratifies calibration by both predicted class and anomaly-score quintiles. This corrects the under-coverage observed in standard Mondrian predictors within high-anomaly operational regimes.

Key Findings and Results
The framework was validated on a held-out set of 289 rare and out-of-taxonomy transients (11 classes) deliberately excluded from training.

• Failure Mode Taxonomy: The authors empirically identify five distinct failure regimes of closed-set classifiers, each requiring different detection signatures:

  1. Continuum-disguised AGN-class events: Confidently misclassified as CVs; only detectable via cross-family non-linear combination.
  2. Spectroscopic disguise within physical neighbours: Rare classes (e.g., Ca-rich SNe, Novae) near-degenerate with training classes; detectable via cross-family combination.
  3. Continuum-dominated mis-classifications: Featureless spectra mapped to wrong classes with intermediate confidence; detectable by distance/density scores.
  4. Legitimately ambiguous progenitor regimes: Genuine physical ambiguity (e.g., FBOTs, ILRTs); detectable via ensemble disagreement and hybrid scores.
  5. Far-out-of-distribution events: Textbook OOD cases; detectable by any sensitive score.
  • Crucial Result: Regimes I and II (constituting ~43% of the rare population) are only recoverable through the cross-family non-linear combination; single-family pipelines fail to detect them.

• Performance:

  • Classification: Achieves 87% overall accuracy and 83% macro recall on the test set.
  • Anomaly Detection: The full cross-family pipeline recovers 82.4% of rare spectra at a 1% false-alarm rate (FAR), compared to ~34% for the base combination without outlier exposure and significantly lower rates for any single-family ensemble.
  • Complementarity: Classifier-internal UQ (e.g., softmax confidence) and embedding-based AD are structurally complementary. Internal UQ dominates on in-distribution boundary errors, while AD dominates on OOD disguise (Regime I). They flag largely disjoint failure populations.

• Operational Capabilities:

  • Enables real-time triage without host redshift or phase.
  • Allows for "AD-orthogonal abstention," where the system can refuse to commit to a prediction based on internal uncertainty while still flagging anomalies based on the AD layer.
  • Provides calibrated risk allocation, allowing operators to set explicit false-alarm budgets.

Significance and Claims
The paper claims to establish the methodological infrastructure for confidence-aware spectroscopic discovery in the Vera C. Rubin Observatory era. Its primary contributions are:

1. First Redshift-Free Multi-Class Classifier: ASTRANet is the first fine-grained, multi-class spectroscopic classifier to operate on observer-frame spectra without requiring redshift or phase at inference, removing a critical real-time dependency.
2. Physics-Driven Failure Taxonomy: It provides the first empirical characterization of five distinct failure modes in spectroscopic classification, demonstrating that a single anomaly-score family is insufficient due to the physical diversity of the anomalies.
3. Structural Complementarity: It establishes that classifier-internal uncertainty and embedding-based anomaly detection are orthogonal axes of confidence that must be deployed simultaneously, not interchangeably.
4. Calibrated Discovery: By integrating conformal prediction (specifically AD-MCP), the framework provides distribution-free guarantees and uniform conditional coverage across anomaly strata, enabling defensible follow-up allocation.

The authors argue that without such a framework, a significant fraction of rare and peculiar transients (specifically those in the "disguised" regimes) will be silently lost to routine classification in future high-volume surveys. The framework is currently being prepared for integration with the ZTF BOOM broker.