Spectral-Influence Augmentation Selection (SIAS)

Principled identification of optimal augmentation strategies for time series foundation model training
cs.LG - Machine Learning

Problem Overview

Time series foundation models rely on data augmentation to extend training coverage, yet augmentation strategies are chosen heuristically before training. SIAS provides a principled method to identify optimal augmentations through a decomposable quality score (affinity + diversity) and online contextual bandit selection.

7 Augmentation Families

Jitter, Scaling, Time Warp, Magnitude Warp, Permutation, Spectral, Trend Injection

4 Domains Tested

Trend (154 trend series), Seasonal (123 seasonal series), Mixed (balanced), Noise (142 AR series)

200 Series per Domain

Length 256, horizon 32, 160 train / 40 validation split, 15 epochs

Augmentation Affinity-Diversity Profiles

Mixed
Trend
Seasonal
Noise

Affinity vs Diversity Scatter

Combined Score Ranking

Insight: Jitter achieves the highest combined score across most domains due to strong diversity with moderate affinity. Scaling preserves structure perfectly (affinity 1.0) but introduces zero spectral diversity.

Training Results

Validation Loss Curves

Trend
Seasonal
Mixed

Final Validation MSE

StrategyTrendSeasonalMixed

Bandit Arm Selection Analysis

Arm Pull Distribution

Trend
Seasonal
Mixed

Average Reward per Arm

Domain-dependent selection: Trend domain: time warp selected 78.7% (118/150 pulls). Seasonal domain: jitter selected 88.0% (132/150 pulls). Mixed domain: jitter selected 82.0% (123/150 pulls).

Spectral Domain Characterization

Average Power Spectral Density

Spectral Context Features

Magnitude Sensitivity Analysis

Jitter: Affinity vs Magnitude

Jitter: Diversity vs Magnitude

Key Findings

1

Decomposable Quality

The affinity-diversity score provides a fast, training-free proxy for augmentation effectiveness. Augmentations ranking high on this score also produce lower validation loss.

2

Domain-Dependent Optimality

The bandit selects different augmentations per domain: time warp for trend data (78.7%), jitter for seasonal data (88.0%).

3

Adaptive Selection Works

SIAS achieves 0.8909 MSE in trend domain, outperforming the best fixed baseline (0.8917) without prior domain knowledge.

4

Degenerate Detection

The framework correctly identifies permutation as destructive (affinity 0.6749, lowest) despite its non-negligible diversity (0.8499).