AMBER

AMBER is an open-source Python library for building, training, and analysing Self-Organizing Maps (SOMs) developed for time series data.
It covers the full workflow — data normalisation, map training, classification, temporal/recurrent SOMs, feature extraction for biosignals, and interactive visualisation.

---

Features

Module	What it provides
Map	Standard SOM with 7 distance metrics, 9 normalisation strategies, 4 weight-initialisation methods, and automatic map-size selection (Vesanto & Alhoniemi 2000)
Classification	BMU assignment, activation map, U-matrix, quantisation error, topological error
TemporalMap	Recurrent SOM (RSOM, Voegtlin 2002) — context vector accumulates temporal structure
TemporalAnalysis	Transition matrix, stability index, mean path length, dwell times
FeatureExtractor	Statistical, spectral, and complexity features for time-series/biosignals
IterativeSOM	Grid-search over map sizes; selects the best map by combined QE+TE score
Visualization	Heat maps, U-matrix, elevation maps, trajectory plots, transition heatmaps

Distance metrics (BMU selection)

euclidean · manhattan · chebyshev · cosine · correlation · dtw · cross_correlation

Normalisation strategies

zscore (fwn) · robust · 01scale · zscore_sample · robust_sample · minmax_sample · l2 (euclidean) · none

---

Installation

From PyPI (recommended)

pip install amber-som

From source

git clone https://github.com/albertonogales/AMBER.git
cd AMBER
pip install -e .

Optional dependencies

Extra	Install command	Required for
librosa	pip install librosa	MFCC feature extraction

---

Quick start

import numpy as np
import AMBER

# --- 1. Train a SOM (map size chosen automatically) ---
data = np.random.rand(200, 8)
som = AMBER.Map(data=data, period=100)
print(f"Map size: {som.map_size}×{som.map_size}")   # e.g. 10×10

# --- 2. Classify ---
cls = AMBER.Classification(som, data)
print(f"Quantisation error : {cls.quantization_error:.4f}")
print(f"Topological error  : {cls.topological_error:.4f}")

# --- 3. Visualise ---
AMBER.Visualization.heat_map(cls)
AMBER.Visualization.umatrix(cls)

Custom map size and distance

som = AMBER.Map(
    data=data,
    size=12,
    period=200,
    distance='dtw',          # DTW distance for BMU selection
    normalization='zscore',
    weights='PCA',
    use_decay=True,
)

Temporal / Recurrent SOM

# Train
tsom = AMBER.TemporalMap(
    data=eeg_windows,
    period=100,
    context_weight=0.5,     # α — context memory
    context_influence=0.3,  # β — context vs. signal balance
    confirm=True,           # required: acknowledge temporal ordering assumption
)

# Analyse temporal dynamics
cls  = AMBER.Classification(tsom, eeg_windows)
ta   = AMBER.TemporalAnalysis(cls)
print(ta.summary())

AMBER.Visualization.trajectory(cls, ta)
AMBER.Visualization.transition_matrix_plot(ta)
AMBER.Visualization.dwell_time_map(ta, cls)

Feature extraction for biosignals

from AMBER import FeatureExtractor

fe = FeatureExtractor(fs=256)            # 256 Hz EEG

# Single window → 1-D feature vector
x = fe.extract(eeg_window, features=[
    'rms', 'zero_crossing_rate',
    'alpha_power', 'spectral_entropy',
    'hjorth_mobility',
])

# Batch of windows → (n_windows, n_features) ready for Map.train
X = fe.extract_batch(eeg_windows, features=['rms', 'alpha_power', 'beta_power'])

---

Application domains

AMBER has been applied to and includes example notebooks for:

Domain	Signal type	Key AMBER feature
EEG / biosignals	Brain, muscle, cardiac	FeatureExtractor (Hjorth, band power) + TemporalMap for state transitions — see eeg_som.ipynb
Audio / speech	MFCC, spectral features	FeatureExtractor(mfcc) + Map for speaker/instrument clustering — see audio_som.ipynb
Climate	SST, MSLP, wind	TemporalMap for weather-regime discovery and transition analysis
Predictive maintenance	Vibration (bearing)	FeatureExtractor (kurtosis, RMS) + degradation trajectory via TemporalAnalysis
Finance / HAR	Returns, accelerometers	TemporalMap for regime / activity detection

---

Running the tests

pip install -r requirements-dev.txt
pytest               # 374 tests, ~99 % coverage

A coverage HTML report is written to coverage_html/.

---

Repository structure

AMBER/
├── AMBER/
│   ├── __init__.py
│   ├── map.py                 # Core SOM
│   ├── classification.py      # BMU assignment & metrics
│   ├── distances.py           # All signal and grid distance functions
│   ├── features.py            # Feature extraction for time series
│   ├── temporal_map.py        # Recurrent SOM (RSOM)
│   ├── temporal_analysis.py   # Temporal metrics
│   ├── iterativesom.py        # Grid-search over map sizes
│   └── visualization.py       # Plots
├── tests/
│   ├── conftest.py
│   ├── test_distances.py
│   ├── test_map.py
│   ├── test_classification.py
│   ├── test_features.py
│   ├── test_temporal_map.py
│   ├── test_temporal_analysis.py
│   ├── test_iterativesom.py
│   ├── test_visualization.py
│   └── test_integration.py
├── examples/
│   ├── colors_example.ipynb              # Basic SOM on RGB colour data
│   ├── eeg_som.ipynb                     # EEG sleep staging (Sleep-EDF)
│   ├── audio_som.ipynb                   # Audio/instrument clustering (NSynth/FSDD)
│   ├── climate_som.ipynb                 # Weather-regime discovery (ERA5)
│   ├── predictive_maintenance_som.ipynb  # Bearing degradation tracking (CWRU)
│   └── ablation_study.ipynb              # Systematic validation of all design choices
├── requirements.txt
├── requirements-dev.txt
└── LICENSE

---

Citing AMBER

If you use AMBER in your research, please cite the companion paper:

@article{nogales2026amber,
  title   = {{AMBER}: Autoassociative Map Builder for tEmporal Representations —
             a Python library for Self-Organizing Maps},
  author  = {Nogales, Alberto and Sicilia, Miguel {\'A}ngel and
             Garc{\'i}a Barriocanal, Elena and Mora Cantallops, Mar{\c{c}}al and
             Ballesteros Rodr{\'i}guez, Alberto},
  journal = {Information Sciences},
  year    = {2026},
  note    = {Under review}
}

A machine-readable CITATION.cff file is included in the repository for automated citation tools (GitHub "Cite this repository", Zenodo, etc.).

---

Contact

• Responsible: Alberto Nogales — alberto.nogales@uah.es
• Supervisors: Alberto Nogales, Miguel Ángel Sicilia

Under license of IERU Research Group at Universidad de Alcalá (Spain)