sublimex 0.1.0

SublimeX: Supervised Bottom-Up Localized Multi-Representative Feature eXtraction for time series and spatial data

SublimeX

Supervised Bottom-Up Localized Multi-Representative Feature eXtraction

SublimeX is an interpretable feature extraction framework for time series and spatial data. It discovers a minimal set of task-specific features through Bayesian optimization, where each feature has explicit, human-readable semantics.

Key Features

• Minimal feature sets: Typically 5-15 features vs. thousands from other methods
• Full interpretability: Each feature = statistic over optimized segment of transformed signal
• Competitive performance: Matches deep learning on many tasks
• Modular design: Custom transforms, objectives, and ML models

Installation

pip install sublimex

Or install from source:

git clone https://github.com/Prgrmmrjns/SublimeX.git
cd SublimeX
pip install -e .

Quick Start

import sublimex
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier

# Your data: list of arrays/DataFrames, one per channel
# Each array has shape (n_samples, n_time_points)
X = [channel1_data, channel2_data, channel3_data]  # 3 channels
y = labels  # Binary or continuous targets

# Split data
X_train, X_test, y_train, y_test = train_test_split(
    range(len(y)), test_size=0.2, stratify=y, random_state=42
)
X_train = [x.iloc[X_train] for x in X]
X_test = [x.iloc[X_test] for x in X]
y_train, y_test = y[X_train], y[X_test]

# Fit SublimeX
model = sublimex.SublimeX(metric='auc', n_trials=100, verbose=True)
train_features = model.fit_transform(X_train, y_train)
test_features = model.transform(X_test)

# Use with any classifier
clf = RandomForestClassifier()
clf.fit(train_features, y_train)
predictions = clf.predict(test_features)

# Interpret discovered features
for desc in model.get_feature_descriptions():
    print(desc)

How It Works

SublimeX discovers discriminative features through a simple but effective process:

1. Signal Transformation: Apply multiple transforms to create different "views" of the data (raw, z-score normalized, derivative, FFT power spectrum)

2. Segment Optimization: Use Bayesian optimization (Optuna) to find segments that maximize downstream model performance

3. Feature Extraction: Compute statistics (e.g., mean) over discovered segments

4. Iterative Discovery: Repeat until adding new features no longer improves performance

Each discovered feature is fully interpretable:

"Mean of z-score normalized signal in channel 2, positions 40-60"

Configuration

Basic Parameters

import sublimex

model = sublimex.SublimeX(
    metric='auc',          # 'auc', 'accuracy', or 'rmse'
    n_trials=300,          # Optimization trials per feature
    inner_cv=1,            # Internal CV folds (1 = single split)
    val_size=0.5,          # Validation size when inner_cv=1
    verbose=True,          # Print progress
)

Custom Transforms

import sublimex

# Register a custom transform
def hilbert_envelope(data):
    from scipy.signal import hilbert
    return np.abs(hilbert(data, axis=-1))

sublimex.register_transform('hilbert', hilbert_envelope)

# Use specific transforms
model = sublimex.SublimeX(
    transforms={
        'raw': lambda x: x,
        'hilbert': hilbert_envelope,
    }
)

Custom Objective Functions

import sublimex
import numpy as np

# Create objective with custom aggregation
def rms(segment):
    """Root mean square."""
    return np.sqrt((segment ** 2).mean(axis=1, keepdims=True))

rms_objective = sublimex.create_custom_objective(rms, 'rms')
model = sublimex.SublimeX(objective_fn=rms_objective)

Custom ML Models

import sublimex
from sklearn.ensemble import RandomForestClassifier

# Wrap any sklearn estimator
rf = RandomForestClassifier(n_estimators=100)
model = sublimex.SublimeX(model=sublimex.SklearnModelWrapper(rf))

Built-in Options

Transforms

Transform	Description	Use Case
raw	Identity (original signal)	Amplitude differences
zscore	Z-score normalization	Shape differences
derivative	First-order gradient	Rate of change, transitions
fft_power	FFT power spectrum	Frequency content, periodicity

Aggregations (for aggregate_objective)

Aggregation	Description
mean	Average value (default)
min, max	Extreme values
range	max - min
std	Standard deviation
median	Robust central tendency
argmin, argmax	Position of extrema

Objectives

Objective	Description
mean_objective	Mean over segment (default, most interpretable)
aggregate_objective	Choose from 8 aggregations
pattern_objective	B-spline pattern matching

Visualization

import sublimex
import matplotlib.pyplot as plt

# Compare feature values across classes
fig = sublimex.plot_feature_distributions(features, y, feature_idx=0)
plt.show()

# Show where a feature's segment falls on the signal
fig = sublimex.plot_segment_on_signal(
    signal, model.extracted_features[0], model.n_time
)
plt.show()

# Compare all transforms for a sample
fig = sublimex.plot_transform_comparison(signal, sublimex.TRANSFORMS)
plt.show()

Saving and Loading

import sublimex

# Save discovered features
model.save_features('features.json')

# Load and reuse
new_model = sublimex.SublimeX()
new_model.load_features('features.json')
features = new_model.transform(X_new)

Complete Example: Synthetic Multi-Channel Time Series

"""Complete example with synthetic data."""
import sublimex
import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score
import lightgbm as lgb

# Generate synthetic multi-channel time series
np.random.seed(42)
n_samples, n_channels, n_bins = 500, 5, 200

X = []
y = np.random.randint(0, 2, n_samples)

for channel in range(n_channels):
    channel_data = []
    for i in range(n_samples):
        t = np.linspace(0, 4 * np.pi, n_bins)
        if y[i] == 1:
            # Class 1: Higher amplitude with distinctive peak
            signal = 2.0 * np.sin(t) + 1.5 * np.sin(2 * t)
            peak = n_bins // 2
            signal[peak-20:peak+20] += 1.5 * np.exp(-((np.arange(40) - 20) ** 2) / 50)
        else:
            # Class 0: Lower amplitude, more noise
            signal = 1.0 * np.sin(t) + 0.5 * np.sin(3 * t)
        signal += np.random.normal(0, 0.3, n_bins)
        channel_data.append(signal)
    
    X.append(pd.DataFrame(channel_data, columns=[f'bin_{j}' for j in range(n_bins)]))

# Split data
idx_train, idx_test = train_test_split(
    range(len(y)), test_size=0.2, stratify=y, random_state=42
)
X_train = [x.iloc[idx_train].astype(np.float32) for x in X]
X_test = [x.iloc[idx_test].astype(np.float32) for x in X]
y_train, y_test = y[idx_train], y[idx_test]

# Discover features with SublimeX
model = sublimex.SublimeX(metric='auc', n_trials=50, verbose=True)
train_features = model.fit_transform(X_train, y_train)
test_features = model.transform(X_test)

print(f"Discovered {train_features.shape[1]} features")

# Train classifier
clf = lgb.LGBMClassifier(n_estimators=100, verbose=-1)
clf.fit(train_features, y_train)
auc = roc_auc_score(y_test, clf.predict_proba(test_features)[:, 1])
print(f"Test AUC: {auc:.4f}")

# Interpret features
print("\nDiscovered Features:")
for desc in model.get_feature_descriptions():
    print(f"  {desc}")

# Visualize
import matplotlib.pyplot as plt

fig = sublimex.plot_feature_importance(
    clf.feature_importances_, 
    model.get_feature_descriptions()
)
plt.show()

Comparison with Other Methods

Method	Features	Interpretability	Optimization
SublimeX	5-15	High (explicit segments)	Bayesian
tsfresh	100-800	Medium (statistical)	Filter
catch22	22	Medium (fixed set)	None
MiniRocket	~10,000	Low	Deterministic
RDST	2k-10k	Medium (shapelets)	Random

Citation

If you use SublimeX in your research, please cite:

@software{sublimex2025,
  title={SublimeX: Supervised Bottom-Up Localized Multi-Representative Feature eXtraction},
  author={Wolber, J.C.},
  year={2026},
  url={https://github.com/Prgrmmrjns/SublimeX}
}

License

MIT License - see LICENSE for details.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.