WEANet: Bridging wavelet inductive bias with network parameter initialization for time series modeling.

Area of Science:

• Deep Learning
• Time Series Analysis
• Signal Processing

Background:

• Traditional deep learning weight initialization methods are general-purpose and ignore time series data structures.
• Effective initialization is crucial for deep neural network performance.

Purpose of the Study:

• Introduce the Wavelet-Enhanced Adaptive Network (WEANet) for principled, domain-aware deep time series modeling.
• Embed wavelet-based inductive bias into network parameters to leverage time series properties.

Main Methods:

• WEANet initializes convolutional kernels with multi-family wavelet coefficients, creating adaptive multi-resolution analyzers.
• A dual-objective loss function with a reconstruction term preserves wavelet-induced representations during training.
• The model was evaluated on classification, forecasting, imputation, and anomaly detection tasks.

Main Results:

• WEANet consistently outperformed state-of-the-art baselines on 24 out of 30 UEA benchmark datasets.
• Achieved up to 5% lower error in long-term forecasting tasks.
• Demonstrated effectiveness as a plug-and-play tokenizer for Transformer models.

Conclusions:

• Wavelet-guided initialization offers a general paradigm for deep time series modeling.
• WEANet successfully integrates signal processing interpretability with deep learning flexibility.
• The proposed method enhances performance and preserves structured representations in time series deep learning.