DyWPE: Signal-Aware Dynamic Wavelet Positional Encoding for Time Series Transformers
DyWPE: Signal-Aware Dynamic Wavelet Positional Encoding for Time Series Transformers
Authors: Habib Irani, Vangelis Metsis Date: 2026-04-21 Paper ID: openalex:2509.14640
Summary
DyWPE proposes a signal-aware positional encoding framework for time series transformers that replaces standard, signal-agnostic index-based encodings with dynamic embeddings derived from the Discrete Wavelet Transform (DWT). By capturing multi-scale non-stationary dynamics, the method addresses a fundamental limitation in applying transformer architectures to complex time-dependent signals. Empirical evaluation across ten datasets shows that DyWPE consistently improves predictive performance, notably on long-range sequences and biomedical data.
Key Contributions
- Introduces DyWPE, a framework that generates positional embeddings for time series using Discrete Wavelet Transform (DWT) to capture multi-scale non-stationary signal dynamics.
- Demonstrates superior performance over traditional index-based positional encodings across ten diverse time series datasets, particularly in long-sequence and biomedical signal tasks.
- Provides a signal-aware alternative to static positional encoding, effectively bridging the gap between temporal sequence indices and underlying signal characteristics.
Open Questions & Future Work
Key Concepts
- dynamic-wavelet-positional-encoding: A positional encoding framework for time series transformers that generates signal-aware embeddings using the Discrete Wavelet Transform.
Archivist Review
I approved the core architectural concept of signal-aware positional encoding using wavelets because it addresses a fundamental limitation in time-series transformers and is highly reusable. I also approved the open question regarding adaptive wavelet decomposition scaling as it represents a significant, non-trivial bottleneck for making such multi-scale methods robust across diverse datasets. Standard practice was applied to reject generic dataset references and routine evaluation claims.
Approved Concepts
- Dynamic Wavelet Positional Encoding: It provides a reusable mechanism to inject signal-aware, non-stationary temporal dynamics into transformer architectures, overcoming the limitations of index-only positional encodings.
Approved Open Questions
- Adaptive Wavelet Decomposition Scaling: Determining the optimal wavelet decomposition depth dynamically would enable more robust and generalized architectures that adapt their hierarchical analysis to the specific complexity of input signals.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2509.14640
- paper_id
- 2509.14640
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- dynamic-wavelet-positional-encoding
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-24T05:50:35Z