How to Sample Effectively in the Big Data Era: Methods and Best Practices

1. Common Sampling Methods

Simple Random Sampling

Method Overview Randomly select samples from the population, giving each individual equal probability.

Algorithm Steps

Assume dataset size N and desire n samples.

Use a random number generator or shuffle algorithm to randomly choose n data points.

Form the final sample set.

Reservoir Sampling

Method Overview Reservoir sampling is suitable for data streams of unknown total size, guaranteeing equal selection probability for all data points.

Algorithm Steps

Maintain a reservoir of size k.

Initialize by filling the reservoir with the first k elements of the stream.

For each subsequent element i (i > k), replace an existing element in the reservoir with probability k/i; otherwise skip.

After processing, the k elements in the reservoir constitute the sample.

Example: In a subway station with unknown daily passenger count, use reservoir sampling to select 100 passengers for a satisfaction survey, ensuring each passenger has an equal chance of being chosen.

Stratified Sampling

Method Overview Applicable when the dataset contains distinct categories or groups, ensuring each category is properly represented in the sample.

Algorithm Steps

Divide the dataset into strata (e.g., gender, age, region).

Perform random sampling within each stratum, either proportionally or with weighted importance.

Oversampling and Undersampling

Method Overview Used for imbalanced class datasets such as fraud detection or medical diagnosis.

Algorithm Steps

Oversampling : Duplicate minority class samples or generate new ones using SMOTE (Synthetic Minority Over‑sampling Technique).

Undersampling : Randomly remove majority class samples to reduce imbalance.

Weighted Sampling

Method Overview Suitable when data points have different importance, e.g., in recommendation systems or financial market analysis.

Algorithm Steps

Compute a weight for each data point based on criteria such as frequency or transaction amount.

Sample according to the weight probabilities rather than uniformly.

2. Sampling Methods Comparison

Simple Random Sampling (SRS) – Advantages: simple concept, easy to implement, unbiased. Disadvantages: high storage and access cost for large datasets, may cause class imbalance. Suitable for market surveys, A/B testing.

Reservoir Sampling – Advantages: works for data streams without storing all data, equal selection probability. Disadvantages: cannot predetermine which data will enter the sample, not suitable for sampling specific categories. Suitable for network log analysis, real‑time monitoring.

Stratified Sampling – Advantages: maintains proportional representation of each class, improves representativeness. Disadvantages: requires prior stratification, adds preprocessing complexity; unsuitable for data without clear categories. Suitable for medical data analysis, customer surveys.

Oversampling – Advantages: addresses class imbalance, improves model learning for minority class, no data loss. Disadvantages: may cause overfitting due to duplicated samples. Suitable for financial fraud detection, medical diagnosis.

Undersampling – Advantages: reduces computational cost, balances data distribution. Disadvantages: possible information loss, affecting overall model performance. Suitable for fraud detection, imbalanced data handling.

Weighted Sampling – Advantages: effective when data points have different importance, enhances analysis efficiency. Disadvantages: requires additional weight information, higher computational complexity. Suitable for recommendation systems, financial market analysis.

In the big‑data era, sampling is not only a way to save computational resources but also a crucial technique for improving data‑analysis quality and decision‑making accuracy.