ProtoX-AD: Self-Explainable Time Series Anomaly Detection and Characterization

Technical Summary: ProtoX-AD

Problem Statement
Time series anomaly detection (TSAD) is critical for domains such as finance, IoT, and healthcare. While recent self-supervised classification-based approaches (SSC-TSAD) have demonstrated strong performance by training classifiers to recognize transformation-specific patterns in normal data, they suffer from a significant lack of explainability. These "black-box" methods provide anomaly scores but offer limited insight into the characteristics of flagged anomalies, failing to support the identification and interpretation of distinct anomalous profiles. Furthermore, existing explainable baselines often lack mechanisms to systematically analyze how transformation design influences both detection performance and the semantic meaning of explanations.

Methodology: ProtoX-AD
The authors propose ProtoX-AD, a prototype-based, self-explainable framework for self-supervised TSAD. The method integrates prototypical self-explainable models (SEMs) with the SSC-TSAD paradigm.

• Architecture: The pipeline comprises five components:

  1. Transformation Module: Applies a set of $K$ transformations (manual or learnable neural networks) to input samples to generate augmented views. The first transformation is an identity function.
  2. Feature Extraction: A probabilistic encoder (inspired by ProtoVAE) maps augmented views into a latent space, modeling the posterior distribution as a tuple of mean and variance.
  3. Dual Reconstruction: Consists of an explainer (reconstructs augmented views from latent representations) and a semantic preservation decoder (reconstructs the original sample). This ensures input-level fidelity and semantic consistency across transformations.
  4. Prototype Module: Learns $M$ prototypes per transformation class in the latent space. These prototypes represent interpretable concepts.
  5. Classification Module: Performs self-supervised classification by computing similarity between latent representations and class-specific prototypes, feeding these similarities into a linear classifier.

• Training Objective: The model is trained on unlabeled normal data using a composite loss function:

  • Classification Loss ($L_{class}$): Cross-entropy loss to distinguish transformation classes.
  • Dual Reconstruction Loss ($L_{recon}$): Includes explainer reconstruction, semantic preservation, and a mixture-based regularization (KL-divergence) to align latent representations with prototypes.
  • Prototype Learning Loss ($L_{proto}$): Combines a clustering loss (ensuring samples are close to at least one prototype) and a coverage loss (preventing "ghost" or unused prototypes).

• Anomaly Detection & Explanation:

  • Detection: The anomaly score is defined as the cross-entropy loss of the identity view's classification. High error indicates the sample violates learned normal patterns.
  • Explanation: For a new sample, the nearest prototype in the latent space is identified and decoded into the input space via the explainer network. This provides a visual, human-interpretable representation of the anomalous profile.

Key Contributions

1. Framework Proposal: Introduction of ProtoX-AD, a self-explainable framework that learns transformation-aware latent representations alongside interpretable prototypes.
2. Performance vs. Explainability: Demonstration that the method achieves detection performance comparable to black-box self-supervised counterparts while providing consistent, semantically meaningful explanations for anomaly characterization.
3. Transformation Analysis: A systematic analysis of how transformation design (manual vs. learnable neural) impacts detection performance, explainability, and the ability to characterize distinct anomalous profiles.

Experimental Results
The authors evaluated ProtoX-AD on three datasets: synthetic (UMD), global temperature anomalies (GTA), and real-world water leak detection (Yorkshire).

• Detection Performance:

  • With manually designed transformations (leveraging domain knowledge), ProtoX-AD achieved performance comparable to or slightly better than black-box self-supervised baselines and consistently outperformed shallow methods (Isolation Forest, LOF, One-Class SVM).
  • With learnable neural transformations, performance dropped significantly across all datasets, particularly in complex scenarios, often underperforming shallow methods.
  • ProtoX-AD generally showed more consistent performance across different datasets and settings compared to the explainable baseline KMEx.

• Explainability Quality:

  • Qualitative: ProtoX-AD produced smoother, class-level prototypes that captured transformation-induced patterns (e.g., specific bell-shaped deviations or temperature shifts). In contrast, KMEx prototypes were sharper and exhibited instance-specific noise. Neural transformations failed to produce meaningful prototypes, often mapping anomalies to the identity (normal) class.
  • Quantitative: ProtoX-AD (with manual transformations) achieved lower Mean Absolute Error (MAE) and Mean Squared Error (MSE) between test samples and their assigned prototype explanations compared to KMEx.
  • Consistency: ProtoX-AD enforced a consistent alignment between latent-space assignments and input-space representations, whereas KMEx exhibited discrepancies where the nearest latent prototype did not correspond to the nearest input-space concept.

Significance and Claims
The paper claims that ProtoX-AD successfully bridges the gap between high-performance anomaly detection and intrinsic explainability. The authors emphasize that incorporating domain knowledge through manually designed transformations is essential for both strong detection performance and meaningful anomaly characterization. While neural transformations offer generality, the results suggest they are insufficient for capturing the semantic structure required for effective self-supervised TSAD and characterization without prior knowledge.

The work concludes that ProtoX-AD offers a more stable detection mechanism than existing explainable baselines by integrating prototypes directly into the learning objective. However, the authors modestly note that the reliance on manual transformations may be a limitation in scenarios where domain knowledge is unavailable, suggesting future work should explore adaptive approaches to progressively learn transformations based on detected anomalies. The framework is presented as extensible to other data modalities beyond time series.