Skip to content

🌊 MultiScaleSeasonMixing

🌊 MultiScaleSeasonMixing

🟡 Intermediate ✅ Stable ⏱️ Time Series

🎯 Overview

The MultiScaleSeasonMixing layer mixes seasonal patterns across multiple time scales in a bottom-up (coarse-to-fine) fashion. It:

  1. Downsamples seasonal patterns to coarser scales
  2. Applies Dense Transformations at each scale
  3. Combines information from multiple scales
  4. Produces Multi-Scale Representations of seasonality

Used as part of TimeMixer's encoder to capture seasonality at different resolutions.

🔍 How It Works

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
Input Seasonal Patterns (Fine Scale)
            |
            V
    +-------------------+
    | Apply Dense       |
    | Transformations   |
    +--------+---------+
             |
             V
    Output Scale 1 (Fine)
             |
             V
    +-------------------+
    | Downsample x2     |
    +--------+---------+
             |
             V
    +-------------------+
    | Apply Dense       |
    | Transformations   |
    +--------+---------+
             |
             V
    Output Scale 2 (Coarser)
             |
             V
            ...

💡 Why Use This Layer?

Challenge Solution
Single Scale ✅ Multi-scale analysis
Loss of Detail ✅ Bottom-up blending
Seasonal Complexity ✅ Hierarchical patterns

📊 Use Cases

  • Multi-Seasonal Data: Multiple overlapping seasonal patterns
  • Hierarchical Forecasting: Predictions at different granularities
  • Pattern Discovery: Seasonal patterns at various scales
  • TimeMixer Encoder: Core component of forecasting model

🚀 Quick Start

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
import keras
from kerasfactory.layers import MultiScaleSeasonMixing

# Create seasonal mixing layer
season_mix = MultiScaleSeasonMixing(
    seq_len=96,
    down_sampling_window=2,
    down_sampling_layers=2
)

# Input: list of seasonal patterns at different scales
x_list = [keras.random.normal((32, 96, 64))]

# Mix across scales
output = season_mix(x_list)
print(len(output))  # Number of output scales

🔧 API Reference

1
2
3
4
5
6
7
kerasfactory.layers.MultiScaleSeasonMixing(
    seq_len: int,
    down_sampling_window: int = 2,
    down_sampling_layers: int = 1,
    name: str | None = None,
    **kwargs: Any
)

Parameters

Parameter Type Default Description
seq_len int — Sequence length
down_sampling_window int 2 Downsampling factor
down_sampling_layers int 1 Number of downsampling layers
name str \| None None Optional layer name

Input

  • List of tensors, each shape (batch, channels, seq_len)

Output

  • List of mixed seasonal patterns at multiple scales

💡 Best Practices

  1. Down-sampling Factor: 2-4 typical values
  2. Number of Layers: 1-3 for most cases
  3. Sequence Length: Must be divisible by downsampling factors
  4. Input Order: Pass finest to coarsest scales

⚠️ Common Pitfalls

  • ❌ Non-divisible seq_len: Causes shape mismatches
  • ❌ Too many layers: Loss of fine-scale information
  • ❌ Wrong input format: List required, not concatenated

📚 References

  • Zhou, T., et al. (2023). "TimeMixer: Decomposing Time Series for Forecasting"

Last Updated: 2025-11-04 | Keras: 3.0+ | Status: ✅ Production Ready