Subsampling

1. General Definition

• Subsampling = selecting a subset of the original dataset (or signal) for analysis or training.
• It can be done for efficiency, class balancing, or signal processing (reducing sample rate).

2. In Machine Learning / Data Science

• Often used when datasets are too large or imbalanced.
• Random subsampling: randomly pick a subset of data (like bootstrapping but without replacement).
• Undersampling (subsampling majority class): reduce the size of the majority class to balance with the minority class.
• Cross-validation subsampling: select subsets of data in each fold for model validation.

Example:

• Dataset: 1,000,000 samples
• You take a subsample of 100,000 to train faster.

3. In Signal Processing / Time Series

• Subsampling = reducing the sampling rate of a signal (a form of downsampling).
• Example:
  • Original: audio sampled at 44.1 kHz
  • Subsampled: reduce to 22.05 kHz
• Must apply a low-pass filter first to avoid aliasing (distortion caused by high frequencies folding into lower ones).

4. Advantages

• Faster training and inference (less data).
• Reduces storage and computation cost.
• In imbalanced datasets, helps balance class proportions (if applied to the majority class).

5. Disadvantages

• Information loss: discards data, which may reduce accuracy.
• If subsampling isn’t stratified, it may change class distribution unintentionally.
• In signals, careless subsampling without filtering introduces aliasing noise.

6. Examples

In ML (Python, scikit-learn):

from sklearn.utils import resample

# Subsample dataset
X_sub, y_sub = resample(X, y, n_samples=10000, random_state=42)

from sklearn.utils import resample

# Subsample dataset
X_sub, y_sub = resample(X, y, n_samples=10000, random_state=42)

In Signal Processing (Python, scipy):

import scipy.signal as sps

# Downsample signal by factor of 2
signal_sub = sps.resample(signal, len(signal)//2)

import scipy.signal as sps

# Downsample signal by factor of 2
signal_sub = sps.resample(signal, len(signal)//2)

7. Comparison

Summary

• Subsampling = selecting a smaller subset of data or reducing signal sampling rate.
• In ML: improves efficiency or balances classes.
• In DSP: reduces sample rate → must use low-pass filtering to avoid aliasing.
• Pros: faster, cheaper. Cons: possible loss of information.