Generalist Representation, Specialist Detection: TS-Router for Time-Series Anomaly Detection
Time-series anomaly detection is hard to generalize across datasets because heterogeneous dynamics favor different anomaly criteria. TS-Router solves this by separating transferable temporal representation—learned by a frozen foundation encoder—from target-specific anomaly decisions coordinated by a lightweight router. The router estimates specialist competence from pretrained representations and selects a small subset of detectors to fit unsupervisedly on each target, requiring no target labels at deployment.
TS-Router reframes foundation models in TSAD as coordinators of specialized criteria rather than universal scorers, backed by strong average ranks across 16 benchmarks. However, the theoretical transfer guarantees rely on strong, unverifiable assumptions about simulated-to-real distribution overlap, and the large gap between the router and the per-target oracle indicates that competence estimation remains noisy.
The core conceptual separation between generalist representation and specialist detection is well motivated and empirically supported. Ablations show that frozen TSFM encoders consistently outperform conventional features for competence estimation, and that adaptive Top-3 routing improves over both the best globally fixed specialist and a full-pool ensemble under all four metrics.
The regret bound assumes absolute continuity and uniformly bounded density ratios between real and simulated representation distributions, along with uniform competence stability—conditions that are impossible to verify in practice and unlikely to hold globally.
Empirically, the oracle-to-router gap remains large (e.g., VUS-PR $68.05$ versus $56.36$ on univariate benchmarks), and the framework ignores cross-variable interactions when routing on multivariate series, limiting expressiveness on benchmarks like SMD and SWaT.
The evidence supports the main claims, though with caveats. TS-Router clearly leads in average rank across benchmarks, and controlled ablations against Best Fixed Top-3 establish that adaptive selection provides real value over any globally fixed set. The mixed-protocol comparison (zero-shot versus target-fitted) is transparently disclosed, but the paper does not rule out that a stronger single target-fitted model could match the routed ensemble.
Reproducibility is mixed. The authors publish an anonymous code link and document hyperparameters and datasets in appendices. However, the Syn-RCD simulator specifications, which are critical to reproducing the routing supervision, are relegated to supplementary materials not shown here; without these details, independently building the competence dataset would be difficult. The choice of routing budget $k=3$ and temperature $\tau$ are also only briefly justified in the main text.
Time-series anomaly detection (TSAD) is difficult to generalize across datasets because heterogeneous temporal dynamics imply different notions of normality and favor different detection criteria. While time-series foundation models provide transferable representations, coupling them with a fixed anomaly-scoring mechanism can overlook this variation. This motivates a different perspective on foundation-model-based TSAD: using foundation models to coordinate specialized anomaly criteria rather than directly imposing a universal one. Based on this view, we propose \textbf{TS-Router}, a generalist-representation, specialist-detection framework that estimates the relative competence of heterogeneous anomaly detectors from pretrained temporal representations and selects suitable specialists for each target series. To avoid relying on specialist-performance labels from real tasks, we derive soft competence supervision from specialists' relative performance on labeled simulated tasks. At deployment, routing requires no target anomaly labels, and only the selected specialists are fitted unsupervisedly on the target series. We bound Top-\(k\) set-competence regret under representation coverage and conditional competence stability. Across 16 real-world benchmarks and four complementary evaluation metrics, TS-Router achieves the best overall average rank. Controlled ablations with multiple frozen TSFM encoders further support the use of pretrained representations for competence estimation and adaptive specialist selection. The code is available at https://anonymous.4open.science/r/TS-Router-D8FF.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.