Skip to content

SeesawNet: Towards Non-stationary Time Series Forecasting with Balanced Modeling of Common and Specific Dependencies

Home / Papers / SeesawNet: Towards Non-stationary Time Series Forecasting with Balanced Modeling of Common and Specific Dependencies

SeesawNet: Towards Non-stationary Time Series Forecasting with Balanced Modeling of Common and Specific Dependencies

Authors: Hao Li, Lu Zhang, Liu Chong, Yankai Chen, Pengyang Wang, Yingjie Zhou Date: 2026-05-14 Paper ID: openalex:2605.14551

Summary

SeesawNet addresses the limitations of instance normalization in non-stationary time series forecasting, which often obscures instance-specific structural information. By introducing Adaptive Stationary-Nonstationary Attention (ASNA), the framework captures common dependencies from normalized data and specific dependencies from raw data, fusing them dynamically based on instance-level characteristics. This dual-stream approach, combined with alternating temporal and channel modeling, effectively reconciles the trade-off between distribution alignment and capturing local dynamics.

Key Contributions

  • Proposed SeesawNet, which balances common and instance-specific dependency modeling to mitigate over-smoothing caused by traditional instance normalization.
  • Introduced Adaptive Stationary-Nonstationary Attention (ASNA) to dynamically fuse patterns from normalized and raw sequences based on instance-specific non-stationarity.
  • Demonstrated consistent performance improvements over existing state-of-the-art models across standard time series forecasting benchmarks.

Open Questions & Future Work

Key Concepts

Archivist Review

The review approved the central attention mechanism as a reusable concept for handling the trade-off between instance normalization and information loss. The open question was reframed to highlight the broader bottleneck of dependency balancing in non-stationary regimes, avoiding generic boilerplate future work. The model architecture itself was rejected as a concept, as it is a specific instantiation of the proposed attention mechanism.

Approved Concepts

  • Adaptive Stationary-Nonstationary Attention (ASNA): It addresses the fundamental tension between modeling common global patterns and instance-specific local dynamics in non-stationary time series, which is the primary contribution of the paper.

Approved Open Questions

  • Non-stationary Pattern Dependency Balance: This is essential for designing more robust forecasting models that do not rely on static assumptions about normalization, as different types of non-stationarity likely require distinct strategies for balancing common and specific signals.

Rejected Candidates

  • [concept] SeesawNet (seesaw-net) - subcomponent_of_broader_mechanism: SeesawNet is the specific model architecture presented; the underlying attention mechanism (ASNA) is the more reusable and central concept.

Metadata & Links

url
https://arxiv.org/abs/2605.14551
paper_id
2605.14551
paper_source
openalex
domain
time-series
tags
time-series-forecastingnon-stationary-time-seriesattention-mechanism
architectures
datasets
concept_slugs
adaptive-stationary-nonstationary-attention
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-17T06:07:11Z