Filtered Conformal Ellipsoids for Graph-Native Time Series

2026-06-15 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Filtered Conformal Ellipsoids introduce a novel method for generating joint prediction sets for multivariate time series, specifically designed to control a single event while adapting to cross-coordinate dependence. This approach utilizes a frozen state-space filter to emit a one-step predictive mean and covariance, applying split-conformal calibration to the resulting Mahalanobis scores. The filter determines the ellipsoid's shape, while conformal calibration sets its scalar radius, leveraging learned predictive covariance without relying on Gaussian tail probabilities for coverage. The theoretical analysis addresses dependencies in filtered scores and non-contracting recurrent filters by analyzing contraction in an observable predictive-law quotient. The framework, instantiated with a GCN-GRU filter using diagonal-plus-low-rank covariance, demonstrated sharper at-target ellipsoids on graph-native traffic benchmarks like METRLA-20 and PEMSBAY-50 compared to static-covariance and non-filter baselines, though factor and copula baselines performed stronger on full-graph scale and non-graph-native datasets.

Key takeaway

For AI Scientists developing robust prediction sets for multivariate graph-native time series, this framework offers a method to achieve valid coverage without strong Gaussian assumptions. You should consider implementing filtered conformal ellipsoids, particularly with GCN-GRU filters, for improved sharpness on datasets like METRLA-20 and PEMSBAY-50. Be aware that factor and copula baselines might be stronger on full-graph scale or non-graph-native data.

Key insights

Filtered conformal ellipsoids provide valid prediction sets for multivariate time series by combining state-space filtering with split-conformal calibration.

Principles

• Conformal calibration can choose scalar radius.
• Learned predictive covariance improves coverage.
• Contraction analysis in predictive-law quotient is key.

Method

A frozen state-space filter emits one-step predictive mean and covariance. Split-conformal calibration is applied to Mahalanobis scores, with the filter shaping the ellipsoid and calibration choosing its radius.

In practice

• Instantiate with GCN-GRU filter.
• Test on graph-native traffic benchmarks.
• Compare against static-covariance baselines.

Topics

• Multivariate Time Series
• Conformal Prediction
• Graph Neural Networks
• State-Space Filters
• Prediction Sets
• Mahalanobis Distance
• GCN-GRU

Best for: AI Scientist, Research Scientist

Related on AIssential

• Filtered Conformal Ellipsoids for Graph-Native Time Series — Filtered conformal ellipsoids use learned covariance for shape and calibration for radius, achieving sharper, valid joint prediction sets.
• Conformal Prediction with Time-Series Data via Sequential Conformalized Density Regions — SCDR offers doubly robust, asymptotically valid conformal prediction for time-series, generating informative, potentially disconnected prediction sets.
• NPMixer: Hierarchical Neighboring Patch Mixing for Time Series Forecasting — NPMixer uses hierarchical patching and adaptive wavelet transforms for superior multivariate time series forecasting.
• GeoGNN: Time Series Geo-Localization using Two-Tower Graph Neural Networks — GeoGNN's two-tower GNN infers time series geographic origins, boosting accuracy by integrating spatial and temporal embeddings.
• Spectral Adaptive Conformal Prediction for Structured Non-Exchangeable Data — Spectral adaptive conformal prediction extends finite-sample coverage to non-exchangeable time series via local spectral similarity and online miscoverage level updates.
• Continuous Cross-Domain Traffic State Prediction via Memory-Augmented Graph Liquid Time-Constant Networks — MA-GLTC uses memory-augmented graph liquid time-constant networks for continuous, fine-grained cross-domain traffic prediction, improving accuracy in data-scarce regions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.