Skip to content

HELIX: Hybrid Encoding with Learnable Identity and Cross-dimensional Synthesis for Time Series Imputation

Home / Papers / HELIX: Hybrid Encoding with Learnable Identity and Cross-dimensional Synthesis for Time Series Imputation

HELIX: Hybrid Encoding with Learnable Identity and Cross-dimensional Synthesis for Time Series Imputation

Authors: Fengming Zhang, Wenjie Du, Huan Zhang, Ke Yu, Shen Qu Date: 2026-05-04 Paper ID: openalex:2605.02278

Summary

HELIX is a hybrid time series imputation framework that addresses the inconsistency of attention-based models by introducing learnable, persistent feature identities. By decoupling static feature semantic properties from dynamic temporal dependencies, the model learns arbitrary cross-feature relationships end-to-end without needing predefined graph structures. Empirical evaluations across five benchmark datasets demonstrate that HELIX achieves state-of-the-art results, effectively aligning learned dependencies with latent physical and semantic structures.

Key Contributions

  • Introduces HELIX, a hybrid attention architecture that replaces dynamic re-discovery of feature relationships with persistent learnable feature identities.
  • Eliminates reliance on predefined graph topology by learning arbitrary feature dependencies end-to-end from temporal co-variation.
  • Establishes state-of-the-art imputation performance across 5 public benchmark datasets and 21 experimental settings.

Key Concepts

  • learnable-feature-identity: A persistent embedding strategy that encodes intrinsic semantic properties of time series features to provide stable anchors for cross-feature dependency modeling.

Archivist Review

I approved ‘Learnable Feature Identity’ as it introduces a reusable inductive bias for time series attention mechanisms, enabling persistent feature-wise semantic modeling. No open questions or datasets were approved as the paper focuses on a specific architecture improvement without explicitly formulating general research bottlenecks or using unique, non-benchmark datasets.

Approved Concepts

  • Learnable Feature Identity: Provides a stable anchor for cross-feature dependencies, addressing the instability of re-discovering relationships at each layer in standard attention mechanisms.

Metadata & Links

url
https://arxiv.org/abs/2605.02278
paper_id
2605.02278
paper_source
openalex
domain
time-series
tags
time-seriesimputationattention-mechanism
architectures
datasets
concept_slugs
learnable-feature-identity
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-07T06:04:01Z