Re^8: Random sampling a variable length file.

[BrowserUk]

This sounds like a positive assertion that there is no correlation based on order. This is quite different from the negative assertion that you don't know whether there are order related correlations.

I disagree. I cannot know if there is any correlation in the ordering of the records, therefore, for my purposes, there is no such correlation.

Put another way, I am constructing an algorithm to process unknown data, and therefore must construct it on the basis that there is no such correlation.

[pajout]

If you assume any correlation in the ordering of the records, why not to pick, for instance, the first 1000 records?

[BrowserUk]

Because there might be (for example) a time-related correlation. If the file in question were a log file running midnight to midnight, it's possible that there might be less activity at either end of the day than in the middle.

Or (for example), the file might have been previously ordered in some way that means the longer records are at the front and shorter at the end.

Whilst it is possible that a random selection picked across the length of any given file might pick all the longest (or shortest), it is far,far less likely, than likely. That's why you pick a random sample.

It is the whole basis of normal distributions, the 68-95-99.7 rule, and the associated statistics.