# Difference between “Sampling” and “Subsampling”?

I just got this question in my mind, because I have seen so many times in the literature that these two words are being used alternatively, Sampling and Subsampling.

What's the main difference between these two words in literature associated with machine learning, specially clustering section? Are they actually the same or there are some significant deviations in specific areas?

Regards ...

A sample is a portion of the population. A subsample is a portion of the sample.

• Thank you for your answer brother, but I don't it will work for me as I can find it on google and it's not so helping at least for me. I mean what's the practical difference between these two terms, especially in clustering. – SANN Aug 14 '18 at 16:32

@TinderForMidgets gave the exact definition for sample and subsample. In clustering practices, some kind of sampling can be implemented to avoid large computations. Usually some of these algorithms require starting points, and again, to reduce computation time, you take a subsample in your sample and assign them as your starting points. So these two perform two different functions in your clustering algorithms. To give completely exhaustive definition:

• Data Set is the entire collection of data to be analyzed. For inferential purposes, this may be treated as having been sampled from a population. All of the data set items will be classified by the process.

• Supersample is a subset of the data set chosen by simple random sampling. In our examples, it is the entire data set, but for larger data sets it will be considerably smaller. All computations prior to the final classification are performed on the supersample. For problems in moderate dimension (up to 50), the supersample will never need to be larger than 100,000–1,000,000 points, since the estimation error in a sample of this size is already too small to matter.

• Sample is one of several ($R_s$) of size $N_s$ chosen by simple random sampling from the supersample. All intensive search operations are conducted in the sample so that the supersample is only used for one iteration from the best solution found in the sample. The sample size $N_s$ should be chosen to be large enough to reflect the essential structure of the data, while being small enough to keep the computations feasible.
• Subsample is one of several ($R_r$) of size $N_r$ chosen by simple random sampling from the sample that is used to begin iterations on the sample. This number should be very small because great diversity in starting points generates diversity in solutions, and increases the chance of finding the best local maximum of the likelihood.

Source